# Deep Learning for Whole-Slide Tissue Histopathology Classification: A Comparative Study in the Identification of Dysplastic and Non-Dysplastic Barrett’s Esophagus

^{1}

^{2}

^{3}

^{4}

^{5}

^{6}

^{*}

## Abstract

**:**

## 1. Background

## 2. Materials and Methods

#### 2.1. Data Collection

#### 2.2. Esophageal Biopsy Datasets

#### 2.3. Deep Learning-Based Feature Representation

#### 2.3.1. Fully Supervised Feature Learning

#### 2.3.2. Unsupervised Feature Learning

#### Codebook Learning

#### WSI Encoding

#### 2.3.3. Weakly Supervised Feature Learning

#### Multiple Instance Learning

#### Expectation-Maximization Model

#### 2.4. Slide-Level Inference

#### 2.5. Feature Importance

## 3. Experimental Evaluation

#### 3.1. Patch Extraction

#### 3.2. Deep Models Architecture

#### 3.3. Experimental Setup

- FS-RF: Image features were extracted using the fully supervised approach, and the random forest was employed for image-level decision fusion;
- FS-SVM: Image features were extracted using the fully supervised approach, and SVM was employed for image-level decision fusion;
- MIL-RF: Image features were extracted using the MIL approach, and the random forest was employed for image-level decision fusion;
- MIL-SVM: Image features were extracted using the MIL approach, and SVM was employed for image-level decision fusion;
- EM-RF: Image features were extracted using the EM approach, and the random forest was employed for image-level decision fusion;
- EM-SVM: Image features were extracted using the EM approach, and SVM was employed for image-level decision fusion;
- KM-RF: Image features were extracted using an unsupervised approach that applies the k-means clustering algorithm to learn codewords. This model employs the random forest for image-level decision fusion;
- KM-SVM: Image features were extracted using an unsupervised approach that applies the k-means clustering algorithm to learn codewords. This model employs SVM for image-level decision fusion;
- GMM-RF: Image features were extracted using an unsupervised approach that applies the GMM clustering algorithm to learn codewords. This model employs the random forest for image-level decision fusion;
- GMM-SVM: Image features were extracted using an unsupervised approach that applies the GMM clustering algorithm to learn codewords. This model employs SVM for image-level decision fusion.

#### 3.4. Classification Results and Statistical Analysis

#### 3.5. Visualization of ROIs on WSIs

## 4. Discussion

## 5. Conclusions

## Author Contributions

## Funding

## Conflicts of Interest

## Abbreviations

BE | Barrett’s Esophagus |

WSI | Whole-Slide Images |

H&E | Hematoxylin and Eosin |

HD-WLE | High-Definition White-Light Endoscopy |

NBI | Narrow Band Imaging |

CNN | Convolutional Neural Networks |

SIFT | Scale-Invariant Feature Transform |

CAE | Convolutional Auto-Encoder |

GMM | Gaussian Mixture Model |

TF | Term Frequencies |

MIL | Multiple Instance Learning |

EM | Expectation-Maximization |

ROI | Regions of Interest |

FS | Fully Supervised |

RF | Random Forest |

KM | k-Means Clustering |

AUC | Area Under the ROC Curve |

PCA | Principle Components Analysis |

## References

- Spechler, S.J. Barrett’s esophagus. N. Engl. J. Med.
**2002**, 346, 836–842. [Google Scholar] [CrossRef] [PubMed] - Shaheen, N.J.; Falk, G.W.; Iyer, P.G.; Gerson, L.B. ACG clinical guideline: Diagnosis and management of Barrett’s esophagus. Am. J. Gastroenterol.
**2016**, 111, 30–50. [Google Scholar] [CrossRef] [PubMed] - Falk, G.W.; Rice, T.W.; Goldblum, J.R.; Richter, J.E. Jumbo biopsy forceps protocol still misses unsuspected cancer in Barrett’s esophagus with high-grade dysplasia. Gastrointest. Endosc.
**1999**, 49, 170–176. [Google Scholar] [CrossRef] - Kariv, R.; Plesec, T.P.; Goldblum, J.R.; Bronner, M.; Oldenburgh, M.; Rice, T.W.; Falk, G.W. The Seattle protocol does not more reliably predict the detection of cancer at the time of esophagectomy than a less intensive surveillance protocol. Clin. Gastroenterol. Hepatol.
**2009**, 7, 653–658. [Google Scholar] [CrossRef] [PubMed] - Wani, S.; Rubenstein, J.H.; Vieth, M.; Bergman, J. Diagnosis and management of low-grade dysplasia in Barrett’s esophagus: Expert review from the Clinical Practice Updates Committee of the American Gastroenterological Association. Gastroenterology
**2016**, 151, 822–835. [Google Scholar] [CrossRef] [PubMed] - Shah, T.; Lippman, R.; Kohli, D.; Mutha, P.; Solomon, S.; Zfass, A. Accuracy of probe-based confocal laser endomicroscopy (pCLE) compared to random biopsies during endoscopic surveillance of Barrett’s esophagus. Endosc. Int. Open
**2018**, 6, E414–E420. [Google Scholar] [CrossRef][Green Version] - Vennalaganti, P.; Kanakadandi, V.; Goldblum, J.R.; Mathur, S.C.; Patil, D.T.; Offerhaus, G.J.; Meijer, S.L.; Vieth, M.; Odze, R.D.; Shreyas, S.; et al. Discordance among pathologists in the United States and Europe in diagnosis of low-grade dysplasia for patients with Barrett’s esophagus. Gastroenterology
**2017**, 152, 564–570. [Google Scholar] [CrossRef][Green Version] - Downs-Kelly, E.; Mendelin, J.E.; Bennett, A.E.; Castilla, E.; Henricks, W.H.; Schoenfield, L.; Skacel, M.; Yerian, L.; Rice, T.W.; Rybicki, L.A.; et al. Poor interobserver agreement in the distinction of high-grade dysplasia and adenocarcinoma in pretreatment Barrett’s esophagus biopsies. Am. J. Gastroenterol.
**2008**, 103, 2333–2340. [Google Scholar] [CrossRef] - Krizhevsky, A.; Sutskever, I.; Hinton, G.E. Imagenet classification with deep convolutional neural networks. In Proceedings of the Advances in Neural Information Processing Systems, Lahe Tahoe, NV, USA, 3–6 December 2012; pp. 1097–1105. [Google Scholar]
- Sparks, R.; Madabhushi, A. Explicit shape descriptors: Novel morphologic features for histopathology classification. Med. Image Anal.
**2013**, 17, 997–1009. [Google Scholar] [CrossRef][Green Version] - Chang, H.; Zhou, Y.; Borowsky, A.; Barner, K.; Spellman, P.; Parvin, B. Stacked predictive sparse decomposition for classification of histology sections. Int. J. Comput. Vis.
**2015**, 113, 3–18. [Google Scholar] [CrossRef][Green Version] - Kandemir, M.; Zhang, C.; Hamprecht, F.A. Empowering multiple instance histopathology cancer diagnosis by cell graphs. In International Conference on Medical Image Computing and Computer-Assisted Intervention; Springer: Cham, Switzerland, 2014; pp. 228–235. [Google Scholar]
- Vu, T.H.; Mousavi, H.S.; Monga, V.; Rao, G.; Rao, U.A. Histopathological image classification using discriminative feature-oriented dictionary learning. IEEE Trans. Med. Imaging
**2015**, 35, 738–751. [Google Scholar] [CrossRef][Green Version] - Cruz-Roa, A.; Basavanhally, A.; González, F.; Gilmore, H.; Feldman, M.; Ganesan, S.; Shih, N.; Tomaszewski, J.; Madabhushi, A. Automatic detection of invasive ductal carcinoma in whole slide images with convolutional neural networks. In Proceedings of the Medical Imaging 2014: Digital Pathology, International Society for Optics and Photonics, San Diego, CA, USA, 15–20 February 2014; Volume 9041, p. 904103. [Google Scholar]
- Wei, J.W.; Tafe, L.J.; Linnik, Y.A.; Vaickus, L.J.; Tomita, N.; Hassanpour, S. Pathologist-level classification of histologic patterns on resected lung adenocarcinoma slides with deep neural networks. Sci. Rep.
**2019**, 9, 1–8. [Google Scholar] [CrossRef] [PubMed][Green Version] - Korbar, B.; Olofson, A.M.; Miraflor, A.P.; Nicka, C.M.; Suriawinata, M.A.; Torresani, L.; Suriawinata, A.A.; Hassanpour, S. Deep learning for classification of colorectal polyps on whole-slide images. J. Pathol. Inform.
**2017**, 8, 30. [Google Scholar] [PubMed] - Cruz-Roa, A.; Gilmore, H.; Basavanhally, A.; Feldman, M.; Ganesan, S.; Shih, N.; Tomaszewski, J.; Madabhushi, A.; González, F. High-throughput adaptive sampling for whole-slide histopathology image analysis (HASHI) via convolutional neural networks: Application to invasive breast cancer detection. PLoS ONE
**2018**, 13, e0196828. [Google Scholar] [CrossRef] [PubMed] - Wang, X.; Chen, H.; Gan, C.; Lin, H.; Dou, Q.; Huang, Q.; Cai, M.; Heng, P.A. Weakly supervised learning for whole slide lung cancer image classification. IEEE Trans. Cybern.
**2018**, 50, 3950–3962. [Google Scholar] [CrossRef] - Sali, R.; Adewole, S.; Ehsan, L.; Denson, L.A.; Kelly, P.; Amadi, B.C.; Holtz, L.; Ali, S.A.; Moore, S.R.; Syed, S.; et al. Hierarchical Deep Convolutional Neural Networks for Multi-category Diagnosis of Gastrointestinal Disorders on Histopathological Images. arXiv
**2020**, arXiv:2005.03868. [Google Scholar] - Sudharshan, P.; Petitjean, C.; Spanhol, F.; Oliveira, L.E.; Heutte, L.; Honeine, P. Multiple instance learning for histopathological breast cancer image classification. Expert Syst. Appl.
**2019**, 117, 103–111. [Google Scholar] [CrossRef] - Campanella, G.; Hanna, M.G.; Geneslaw, L.; Miraflor, A.; Silva, V.W.K.; Busam, K.J.; Brogi, E.; Reuter, V.E.; Klimstra, D.S.; Fuchs, T.J. Clinical-grade computational pathology using weakly supervised deep learning on whole slide images. Nat. Med.
**2019**, 25, 1301–1309. [Google Scholar] [CrossRef] - Hou, L.; Samaras, D.; Kurc, T.M.; Gao, Y.; Davis, J.E.; Saltz, J.H. Patch-based convolutional neural network for whole slide tissue image classification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 2424–2433. [Google Scholar]
- Rony, J.; Belharbi, S.; Dolz, J.; Ayed, I.B.; McCaffrey, L.; Granger, E. Deep weakly-supervised learning methods for classification and localization in histology images: A survey. arXiv
**2019**, arXiv:1909.03354. [Google Scholar] - Sali, R.; Ehsan, L.; Kowsari, K.; Khan, M.; Moskaluk, C.; Syed, S.; Brown, D.E. Celiacnet: Celiac disease severity diagnosis on duodenal histopathological images using deep residual networks. In Proceedings of the IEEE International Conference on Bioinformatics and Biomedicine (BIBM 2019), San Diego, CA, USA, 18–21 November 2019; pp. 962–967. [Google Scholar]
- Yamamoto, Y.; Tsuzuki, T.; Akatsuka, J.; Ueki, M.; Morikawa, H.; Numata, Y.; Takahara, T.; Tsuyuki, T.; Tsutsumi, K.; Nakazawa, R.; et al. Automated acquisition of explainable knowledge from unannotated histopathology images. Nat. Commun.
**2019**, 10, 1–9. [Google Scholar] [CrossRef][Green Version] - Tellez, D.; Litjens, G.; van der Laak, J.; Ciompi, F. Neural image compression for gigapixel histopathology image analysis. IEEE Trans. Pattern Anal. Mach. Intell.
**2019**. [Google Scholar] [CrossRef] [PubMed][Green Version] - Caicedo, J.C.; Cruz, A.; Gonzalez, F.A. Histopathology image classification using bag of features and kernel functions. In Conference on Artificial Intelligence in Medicine in Europe; Springer: Berlin/Heidelberg, Germany, 2009; pp. 126–135. [Google Scholar]
- Vanegas, J.A.; Arevalo, J.; González, F.A. Unsupervised feature learning for content-based histopathology image retrieval. In Proceedings of the 2014 12th International Workshop on Content-Based Multimedia Indexing (CBMI), Klagenfurt, Austria, 18–20 June 2014; pp. 1–6. [Google Scholar]
- Alirezazadeh, P.; Hejrati, B.; Monsef-Esfahani, A.; Fathi, A. Representation learning-based unsupervised domain adaptation for classification of breast cancer histopathology images. Biocybern. Biomed. Eng.
**2018**, 38, 671–683. [Google Scholar] [CrossRef] - Popovici, V.; Budinská, E.; Čápková, L.; Schwarz, D.; Dušek, L.; Feit, J.; Jaggi, R. Joint analysis of histopathology image features and gene expression in breast cancer. BMC Bioinform.
**2016**, 17, 209. [Google Scholar] [CrossRef] [PubMed][Green Version] - Xu, Y.; Mo, T.; Feng, Q.; Zhong, P.; Lai, M.; Eric, I.; Chang, C. Deep learning of feature representation with multiple instance learning for medical image analysis. In Proceedings of the 2014 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Florence, Italy, 4–9 May 2014; pp. 1626–1630. [Google Scholar]
- Amores, J. Multiple instance classification: Review, taxonomy and comparative study. Artif. Intell.
**2013**, 201, 81–105. [Google Scholar] [CrossRef] - Available online: https://nanozoomer.hamamatsu.com/jp/en/product/search/C13220-01/index.html (accessed on 10 August 2020).
- Bishop, C.M. Pattern Recognition and Machine Learning; Springer: New York, NY, USA, 2006. [Google Scholar]
- Breiman, L. Random forests. Mach. Learn.
**2001**, 45, 5–32. [Google Scholar] [CrossRef][Green Version] - Vahadane, A.; Peng, T.; Sethi, A.; Albarqouni, S.; Wang, L.; Baust, M.; Steiger, K.; Schlitter, A.M.; Esposito, I.; Navab, N. Structure-preserving color normalization and sparse stain separation for histological images. IEEE Trans. Med. Imaging
**2016**, 35, 1962–1971. [Google Scholar] [CrossRef] [PubMed] - Kowsari, K.; Sali, R.; Khan, M.N.; Adorno, W.; Ali, S.A.; Moore, S.R.; Amadi, B.C.; Kelly, P.; Syed, S.; Brown, D.E. Diagnosis of celiac disease and environmental enteropathy on biopsy images using color balancing on convolutional neural networks. In Proceedings of the Future Technologies Conference; Springer: Cham, Switzerland, 2019; pp. 750–765. [Google Scholar]
- He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 770–778. [Google Scholar]

**Figure 1.**An example of the annotation process on a typical whole-slide image (WSI). Red, green, and yellow highlighted areas indicate areas that were annotated and from which labeled patches were taken. Squamous tissue (green arrowhead), non-dysplastic Barrett’s with Goblet cells (yellow arrowhead), and dysplastic tissue with crowding and hyperchromasia (lower zoomed section) were all present within the same whole-slide image.

**Figure 2.**Overview of the fully supervised feature extraction framework. (

**A**) A convolutional neural network (CNN) is trained on high-resolution tissue tiles sampled from annotated regions of WSIs in the training set. (

**B**) Next, the trained model is employed to output the class’ probability distributions for each high-resolution tissue tile generated from new WSIs. The patch-level probabilities corresponding to all patches derived from a WSI are aggregated into WSI-level probabilities histogram.

**Figure 3.**Overview of the unsupervised feature extraction framework. (

**A**) This is the codebook learning phase. In this step, tissue tiles are sampled from WSIs in the training set. Neither the annotated areas nor the WSIs labels are used in this framework. An encoder is trained in an unsupervised fashion to map each high-resolution tissue tile into a low-dimensional embedding space, and then the Gaussian mixture model (GMM) is employed to cluster extracted features from tissue tiles into a number of clusters. Each cluster is indeed a morphological feature called a codeblock. The set of all codeblocks is called codebook. (

**B**) This is the WSI encoding phase. In this phase, the trained convolutional autoencoder (CAE) is employed to extract embedding features from high-resolution tissue tiles derived from new WSIs. Then, the posterior probabilities of clusters constructed in the previous phase for patch-level extracted features are calculated. Finally, the posterior probabilities corresponding to all patches derived from a WSI are aggregated into a WSI-level probabilities histogram.

**Figure 4.**Overview of weakly supervised feature extraction framework. (

**A**) A CNN is trained on high-resolution tissue tiles sampled from the labeled WSIs in the training set. This model uses only the reported diagnoses as labels for training WSIs and assumes that sampled tissue tiles have the same labels as their corresponding WSIs. (

**B**) Once the training concludes, the trained model is employed to output the class’ probability distributions for each high-resolution tissue tile generated from new WSIs. The patch-level probabilities corresponding to all patches derived from a WSI are aggregated into a WSI-level probabilities histogram.

**Figure 5.**Principle component analysis (PCA) plot for WSIs encoded using (

**A**) fully supervised, (

**B**) multiple instance learning (MIL), (

**C**) expectation-maximization (EM), (

**D**) unsupervised (k-means), and (

**E**) unsupervised (GMM) approaches.

**Figure 6.**Boxplots of the 10-fold cross-validation results for weighted (

**A**) accuracy, (

**B**) area under the ROC curve (AUC), (

**C**) precision, (

**D**) recall, and (

**E**) F1 score in different models.

**Figure 7.**Heatmaps generated by different feature extraction approaches for some samples from (

**A**) dysplastic Barrett’s esophagus (BE) and (

**B**) non-dysplastic BE. Area of attention is shown in red.

Feature Extraction Approach | Clustering Algorithm | Classification Algorithm | Model |
---|---|---|---|

fully supervised | - | Random Forest | FS-RF |

fully supervised | - | SVM | FS-SVM |

weakly supervised (MIL) | - | Random Forest | MIL-RF |

weakly supervised (MIL) | - | SVM | MIL-SVM |

weakly supervised (EM) | - | Random Forest | EM-RF |

weakly supervised (EM) | - | SVM | EM-SVM |

unsupervised | k-means | Random Forest | KM-RF |

unsupervised | k-means | SVM | KM-SVM |

unsupervised | GMM | Random Forest | GMM-RF |

unsupervised | GMM | SVM | GMM-SVM |

Models | Classes | Metrices | ||||
---|---|---|---|---|---|---|

Accuracy | AUC | Precision | Recall | F1 score | ||

FS-RF | Dysplastic BE | $0.563\phantom{\rule{3.33333pt}{0ex}}[0.454,\phantom{\rule{3.33333pt}{0ex}}0.671]$ | $0.616\phantom{\rule{3.33333pt}{0ex}}[0.489,\phantom{\rule{3.33333pt}{0ex}}0.743]$ | $0.395\phantom{\rule{3.33333pt}{0ex}}[0.211,\phantom{\rule{3.33333pt}{0ex}}0.578]$ | $0.451\phantom{\rule{3.33333pt}{0ex}}[0.319,\phantom{\rule{3.33333pt}{0ex}}0.584]$ | $0.367\phantom{\rule{3.33333pt}{0ex}}[0.261,\phantom{\rule{3.33333pt}{0ex}}0.472]$ |

Non-dysplastic BE | $0.655\phantom{\rule{3.33333pt}{0ex}}[0.563,\phantom{\rule{3.33333pt}{0ex}}0.746]$ | $0.758\phantom{\rule{3.33333pt}{0ex}}[0.653,\phantom{\rule{3.33333pt}{0ex}}0.863]$ | $0.565\phantom{\rule{3.33333pt}{0ex}}[0.360,\phantom{\rule{3.33333pt}{0ex}}0.770]$ | $0.555\phantom{\rule{3.33333pt}{0ex}}[0.384,\phantom{\rule{3.33333pt}{0ex}}0.726]$ | $0.520\phantom{\rule{3.33333pt}{0ex}}[0.360,\phantom{\rule{3.33333pt}{0ex}}0.680]$ | |

Squamous | $0.857\phantom{\rule{3.33333pt}{0ex}}[0.755,\phantom{\rule{3.33333pt}{0ex}}0.958]$ | $0.934\phantom{\rule{3.33333pt}{0ex}}[0.866,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.760\phantom{\rule{3.33333pt}{0ex}}[0.565,\phantom{\rule{3.33333pt}{0ex}}0.954]$ | $0.732\phantom{\rule{3.33333pt}{0ex}}[0.566,\phantom{\rule{3.33333pt}{0ex}}0.898]$ | $0.720\phantom{\rule{3.33333pt}{0ex}}[0.558,\phantom{\rule{3.33333pt}{0ex}}0.881]$ | |

Weighted Average | $0.655\phantom{\rule{3.33333pt}{0ex}}[0.560,\phantom{\rule{3.33333pt}{0ex}}0.751]$ | $0.761\phantom{\rule{3.33333pt}{0ex}}[0.647,\phantom{\rule{3.33333pt}{0ex}}0.876]$ | $0.655\phantom{\rule{3.33333pt}{0ex}}[0.507,\phantom{\rule{3.33333pt}{0ex}}0.802]$ | $0.537\phantom{\rule{3.33333pt}{0ex}}[0.416,\phantom{\rule{3.33333pt}{0ex}}0.657]$ | $0.554\phantom{\rule{3.33333pt}{0ex}}[0.430,\phantom{\rule{3.33333pt}{0ex}}0.678]$ | |

FS-SVM | Dysplastic BE | $0.619\phantom{\rule{3.33333pt}{0ex}}[0.480,\phantom{\rule{3.33333pt}{0ex}}0.758]$ | $0.661\phantom{\rule{3.33333pt}{0ex}}[0.481,\phantom{\rule{3.33333pt}{0ex}}0.840]$ | $0.434\phantom{\rule{3.33333pt}{0ex}}[0.246,\phantom{\rule{3.33333pt}{0ex}}0.622]$ | $0.565\phantom{\rule{3.33333pt}{0ex}}[0.379,\phantom{\rule{3.33333pt}{0ex}}0.751]$ | $0.455\phantom{\rule{3.33333pt}{0ex}}[0.303,\phantom{\rule{3.33333pt}{0ex}}0.607]$ |

Non-dysplastic BE | $0.663\phantom{\rule{3.33333pt}{0ex}}[0.551,\phantom{\rule{3.33333pt}{0ex}}0.775]$ | $0.778\phantom{\rule{3.33333pt}{0ex}}[0.665,\phantom{\rule{3.33333pt}{0ex}}0.892]$ | $0.542\phantom{\rule{3.33333pt}{0ex}}[0.323,\phantom{\rule{3.33333pt}{0ex}}0.762]$ | $0.418\phantom{\rule{3.33333pt}{0ex}}[0.162,\phantom{\rule{3.33333pt}{0ex}}0.673]$ | $0.424\phantom{\rule{3.33333pt}{0ex}}[0.217,\phantom{\rule{3.33333pt}{0ex}}0.632]$ | |

Squamous | $0.888\phantom{\rule{3.33333pt}{0ex}}[0.811,\phantom{\rule{3.33333pt}{0ex}}0.964]$ | $0.921\phantom{\rule{3.33333pt}{0ex}}[0.842,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.741\phantom{\rule{3.33333pt}{0ex}}[0.565,\phantom{\rule{3.33333pt}{0ex}}0.916]$ | $0.860\phantom{\rule{3.33333pt}{0ex}}[0.710,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.779\phantom{\rule{3.33333pt}{0ex}}[0.634,\phantom{\rule{3.33333pt}{0ex}}0.923]$ | |

Weighted Average | $0.689\phantom{\rule{3.33333pt}{0ex}}[0.593,\phantom{\rule{3.33333pt}{0ex}}0.784]$ | $0.773\phantom{\rule{3.33333pt}{0ex}}[0.644,\phantom{\rule{3.33333pt}{0ex}}0.902]$ | $0.637\phantom{\rule{3.33333pt}{0ex}}[0.466,\phantom{\rule{3.33333pt}{0ex}}0.807]$ | $0.585\phantom{\rule{3.33333pt}{0ex}}[0.448,\phantom{\rule{3.33333pt}{0ex}}0.721]$ | $0.572\phantom{\rule{3.33333pt}{0ex}}[0.423,\phantom{\rule{3.33333pt}{0ex}}0.721]$ | |

MIL-RF | Dysplastic BE | $0.842\phantom{\rule{3.33333pt}{0ex}}[0.739,\phantom{\rule{3.33333pt}{0ex}}0.945]$ | $0.924\phantom{\rule{3.33333pt}{0ex}}[0.835,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.756\phantom{\rule{3.33333pt}{0ex}}[0.584,\phantom{\rule{3.33333pt}{0ex}}0.929]$ | $0.761\phantom{\rule{3.33333pt}{0ex}}[0.526,\phantom{\rule{3.33333pt}{0ex}}0.996]$ | $0.694\phantom{\rule{3.33333pt}{0ex}}[0.524,\phantom{\rule{3.33333pt}{0ex}}0.863]$ |

Non-dysplastic BE | $0.844\phantom{\rule{3.33333pt}{0ex}}[0.742,\phantom{\rule{3.33333pt}{0ex}}0.945]$ | $0.926\phantom{\rule{3.33333pt}{0ex}}[0.835,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.816\phantom{\rule{3.33333pt}{0ex}}[0.657,\phantom{\rule{3.33333pt}{0ex}}0.976]$ | $0.820\phantom{\rule{3.33333pt}{0ex}}[0.710,\phantom{\rule{3.33333pt}{0ex}}0.930]$ | $0.793\phantom{\rule{3.33333pt}{0ex}}[0.692,\phantom{\rule{3.33333pt}{0ex}}0.893]$ | |

Squamous | $0.990\phantom{\rule{3.33333pt}{0ex}}[0.978,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $1.000\phantom{\rule{3.33333pt}{0ex}}[0.999,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.964\phantom{\rule{3.33333pt}{0ex}}[0.899,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.984\phantom{\rule{3.33333pt}{0ex}}[0.955,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.971\phantom{\rule{3.33333pt}{0ex}}[0.933,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | |

Weighted Average | $0.874\phantom{\rule{3.33333pt}{0ex}}[0.778,\phantom{\rule{3.33333pt}{0ex}}0.971]$ | $0.939\phantom{\rule{3.33333pt}{0ex}}[0.856,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.877\phantom{\rule{3.33333pt}{0ex}}[0.781,\phantom{\rule{3.33333pt}{0ex}}0.974]$ | $0.838\phantom{\rule{3.33333pt}{0ex}}[0.738,\phantom{\rule{3.33333pt}{0ex}}0.938]$ | $0.831\phantom{\rule{3.33333pt}{0ex}}[0.716,\phantom{\rule{3.33333pt}{0ex}}0.947]$ | |

MIL-SVM | Dysplastic BE | $0.845\phantom{\rule{3.33333pt}{0ex}}[0.736,\phantom{\rule{3.33333pt}{0ex}}0.954]$ | $0.918\phantom{\rule{3.33333pt}{0ex}}[0.828,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.744\phantom{\rule{3.33333pt}{0ex}}[0.592,\phantom{\rule{3.33333pt}{0ex}}0.895]$ | $0.775\phantom{\rule{3.33333pt}{0ex}}[0.535,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.707\phantom{\rule{3.33333pt}{0ex}}[0.534,\phantom{\rule{3.33333pt}{0ex}}0.880]$ |

Non-dysplastic BE | $0.847\phantom{\rule{3.33333pt}{0ex}}[0.738,\phantom{\rule{3.33333pt}{0ex}}0.956]$ | $0.935\phantom{\rule{3.33333pt}{0ex}}[0.858,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.840\phantom{\rule{3.33333pt}{0ex}}[0.689,\phantom{\rule{3.33333pt}{0ex}}0.991]$ | $0.808\phantom{\rule{3.33333pt}{0ex}}[0.696,\phantom{\rule{3.33333pt}{0ex}}0.920]$ | $0.799\phantom{\rule{3.33333pt}{0ex}}[0.694,\phantom{\rule{3.33333pt}{0ex}}0.905]$ | |

Squamous | $0.989\phantom{\rule{3.33333pt}{0ex}}[0.978,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.998\phantom{\rule{3.33333pt}{0ex}}[0.995,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.983\phantom{\rule{3.33333pt}{0ex}}[0.946,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.968\phantom{\rule{3.33333pt}{0ex}}[0.931,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.974\phantom{\rule{3.33333pt}{0ex}}[0.949,\phantom{\rule{3.33333pt}{0ex}}0.999]$ | |

Weighted Average | $0.876\phantom{\rule{3.33333pt}{0ex}}[0.774,\phantom{\rule{3.33333pt}{0ex}}0.978]$ | $0.938\phantom{\rule{3.33333pt}{0ex}}[0.862,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.868\phantom{\rule{3.33333pt}{0ex}}[0.758,\phantom{\rule{3.33333pt}{0ex}}0.977]$ | $0.841\phantom{\rule{3.33333pt}{0ex}}[0.733,\phantom{\rule{3.33333pt}{0ex}}0.948]$ | $0.831\phantom{\rule{3.33333pt}{0ex}}[0.704,\phantom{\rule{3.33333pt}{0ex}}0.957]$ | |

EM-RF | Dysplastic BE | $0.837\phantom{\rule{3.33333pt}{0ex}}[0.719,\phantom{\rule{3.33333pt}{0ex}}0.954]$ | $0.896\phantom{\rule{3.33333pt}{0ex}}[0.783,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.668\phantom{\rule{3.33333pt}{0ex}}[0.436,\phantom{\rule{3.33333pt}{0ex}}0.901]$ | $0.747\phantom{\rule{3.33333pt}{0ex}}[0.480,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.676\phantom{\rule{3.33333pt}{0ex}}[0.439,\phantom{\rule{3.33333pt}{0ex}}0.912]$ |

Non-dysplastic BE | $0.836\phantom{\rule{3.33333pt}{0ex}}[0.717,\phantom{\rule{3.33333pt}{0ex}}0.954]$ | $0.915\phantom{\rule{3.33333pt}{0ex}}[0.818,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.809\phantom{\rule{3.33333pt}{0ex}}[0.649,\phantom{\rule{3.33333pt}{0ex}}0.969]$ | $0.817\phantom{\rule{3.33333pt}{0ex}}[0.676,\phantom{\rule{3.33333pt}{0ex}}0.958]$ | $0.794\phantom{\rule{3.33333pt}{0ex}}[0.661,\phantom{\rule{3.33333pt}{0ex}}0.928]$ | |

Squamous | $0.985\phantom{\rule{3.33333pt}{0ex}}[0.974,\phantom{\rule{3.33333pt}{0ex}}0.996]$ | $0.998\phantom{\rule{3.33333pt}{0ex}}[0.994,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.958\phantom{\rule{3.33333pt}{0ex}}[0.893,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.963\phantom{\rule{3.33333pt}{0ex}}[0.925,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.957\phantom{\rule{3.33333pt}{0ex}}[0.918,\phantom{\rule{3.33333pt}{0ex}}0.996]$ | |

Weighted Average | $0.865\phantom{\rule{3.33333pt}{0ex}}[0.760,\phantom{\rule{3.33333pt}{0ex}}0.971]$ | $0.923\phantom{\rule{3.33333pt}{0ex}}[0.825,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.828\phantom{\rule{3.33333pt}{0ex}}[0.670,\phantom{\rule{3.33333pt}{0ex}}0.986]$ | $0.829\phantom{\rule{3.33333pt}{0ex}}[0.713,\phantom{\rule{3.33333pt}{0ex}}0.944]$ | $0.814\phantom{\rule{3.33333pt}{0ex}}[0.671,\phantom{\rule{3.33333pt}{0ex}}0.958]$ | |

EM-SVM | Dysplastic BE | $0.858\phantom{\rule{3.33333pt}{0ex}}[0.738,\phantom{\rule{3.33333pt}{0ex}}0.979]$ | $0.900\phantom{\rule{3.33333pt}{0ex}}[0.805,\phantom{\rule{3.33333pt}{0ex}}0.994]$ | $0.757\phantom{\rule{3.33333pt}{0ex}}[0.524,\phantom{\rule{3.33333pt}{0ex}}0.990]$ | $0.757\phantom{\rule{3.33333pt}{0ex}}[0.486,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.709\phantom{\rule{3.33333pt}{0ex}}[0.474,\phantom{\rule{3.33333pt}{0ex}}0.945]$ |

Non-dysplastic BE | $0.859\phantom{\rule{3.33333pt}{0ex}}[0.739,\phantom{\rule{3.33333pt}{0ex}}0.980]$ | $0.938\phantom{\rule{3.33333pt}{0ex}}[0.870,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.846\phantom{\rule{3.33333pt}{0ex}}[0.688,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.863\phantom{\rule{3.33333pt}{0ex}}[0.733,\phantom{\rule{3.33333pt}{0ex}}0.994]$ | $0.834\phantom{\rule{3.33333pt}{0ex}}[0.705,\phantom{\rule{3.33333pt}{0ex}}0.963]$ | |

Squamous | $0.983\phantom{\rule{3.33333pt}{0ex}}[0.970,\phantom{\rule{3.33333pt}{0ex}}0.996]$ | $1.000\phantom{\rule{3.33333pt}{0ex}}[1.000,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.958\phantom{\rule{3.33333pt}{0ex}}[0.894,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.964\phantom{\rule{3.33333pt}{0ex}}[0.922,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.957\phantom{\rule{3.33333pt}{0ex}}[0.920,\phantom{\rule{3.33333pt}{0ex}}0.995]$ | |

Weighted Average | $0.883\phantom{\rule{3.33333pt}{0ex}}[0.775,\phantom{\rule{3.33333pt}{0ex}}0.991]$ | $0.935\phantom{\rule{3.33333pt}{0ex}}[0.861,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.852\phantom{\rule{3.33333pt}{0ex}}[0.691,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.850\phantom{\rule{3.33333pt}{0ex}}[0.732,\phantom{\rule{3.33333pt}{0ex}}0.969]$ | $0.834\phantom{\rule{3.33333pt}{0ex}}[0.687,\phantom{\rule{3.33333pt}{0ex}}0.981]$ | |

KM-RF | Dysplastic BE | $0.660\phantom{\rule{3.33333pt}{0ex}}[0.488,\phantom{\rule{3.33333pt}{0ex}}0.831]$ | $0.778\phantom{\rule{3.33333pt}{0ex}}[0.651,\phantom{\rule{3.33333pt}{0ex}}0.905]$ | $0.517\phantom{\rule{3.33333pt}{0ex}}[0.238,\phantom{\rule{3.33333pt}{0ex}}0.796]$ | $0.516\phantom{\rule{3.33333pt}{0ex}}[0.232,\phantom{\rule{3.33333pt}{0ex}}0.801]$ | $0.449\phantom{\rule{3.33333pt}{0ex}}[0.212,\phantom{\rule{3.33333pt}{0ex}}0.686]$ |

Non-dysplastic BE | $0.682\phantom{\rule{3.33333pt}{0ex}}[0.522,\phantom{\rule{3.33333pt}{0ex}}0.843]$ | $0.793\phantom{\rule{3.33333pt}{0ex}}[0.642,\phantom{\rule{3.33333pt}{0ex}}0.943]$ | $0.665\phantom{\rule{3.33333pt}{0ex}}[0.440,\phantom{\rule{3.33333pt}{0ex}}0.890]$ | $0.727\phantom{\rule{3.33333pt}{0ex}}[0.539,\phantom{\rule{3.33333pt}{0ex}}0.915]$ | $0.626\phantom{\rule{3.33333pt}{0ex}}[0.440,\phantom{\rule{3.33333pt}{0ex}}0.812]$ | |

Squamous | $0.954\phantom{\rule{3.33333pt}{0ex}}[0.910,\phantom{\rule{3.33333pt}{0ex}}0.998]$ | $0.996\phantom{\rule{3.33333pt}{0ex}}[0.990,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.877\phantom{\rule{3.33333pt}{0ex}}[0.725,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.976\phantom{\rule{3.33333pt}{0ex}}[0.946,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.907\phantom{\rule{3.33333pt}{0ex}}[0.806,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | |

Weighted Average | $0.720\phantom{\rule{3.33333pt}{0ex}}[0.575,\phantom{\rule{3.33333pt}{0ex}}0.865]$ | $0.836\phantom{\rule{3.33333pt}{0ex}}[0.715,\phantom{\rule{3.33333pt}{0ex}}0.957]$ | $0.728\phantom{\rule{3.33333pt}{0ex}}[0.547,\phantom{\rule{3.33333pt}{0ex}}0.908]$ | $0.648\phantom{\rule{3.33333pt}{0ex}}[0.484,\phantom{\rule{3.33333pt}{0ex}}0.812]$ | $0.631\phantom{\rule{3.33333pt}{0ex}}[0.453,\phantom{\rule{3.33333pt}{0ex}}0.810]$ | |

KM-SVM | Dysplastic BE | $0.676\phantom{\rule{3.33333pt}{0ex}}[0.542,\phantom{\rule{3.33333pt}{0ex}}0.809]$ | $0.754\phantom{\rule{3.33333pt}{0ex}}[0.642,\phantom{\rule{3.33333pt}{0ex}}0.867]$ | $0.512\phantom{\rule{3.33333pt}{0ex}}[0.322,\phantom{\rule{3.33333pt}{0ex}}0.702]$ | $0.632\phantom{\rule{3.33333pt}{0ex}}[0.412,\phantom{\rule{3.33333pt}{0ex}}0.852]$ | $0.509\phantom{\rule{3.33333pt}{0ex}}[0.353,\phantom{\rule{3.33333pt}{0ex}}0.662]$ |

Non-dysplastic BE | $0.705\phantom{\rule{3.33333pt}{0ex}}[0.565,\phantom{\rule{3.33333pt}{0ex}}0.845]$ | $0.776\phantom{\rule{3.33333pt}{0ex}}[0.641,\phantom{\rule{3.33333pt}{0ex}}0.912]$ | $0.672\phantom{\rule{3.33333pt}{0ex}}[0.468,\phantom{\rule{3.33333pt}{0ex}}0.876]$ | $0.643\phantom{\rule{3.33333pt}{0ex}}[0.499,\phantom{\rule{3.33333pt}{0ex}}0.788]$ | $0.615\phantom{\rule{3.33333pt}{0ex}}[0.461,\phantom{\rule{3.33333pt}{0ex}}0.769]$ | |

Squamous | $0.939\phantom{\rule{3.33333pt}{0ex}}[0.884,\phantom{\rule{3.33333pt}{0ex}}0.995]$ | $0.976\phantom{\rule{3.33333pt}{0ex}}[0.935,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.852\phantom{\rule{3.33333pt}{0ex}}[0.695,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.946\phantom{\rule{3.33333pt}{0ex}}[0.883,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.880\phantom{\rule{3.33333pt}{0ex}}[0.768,\phantom{\rule{3.33333pt}{0ex}}0.991]$ | |

Weighted Average | $0.733\phantom{\rule{3.33333pt}{0ex}}[0.608,\phantom{\rule{3.33333pt}{0ex}}0.857]$ | $0.812\phantom{\rule{3.33333pt}{0ex}}[0.695,\phantom{\rule{3.33333pt}{0ex}}0.929]$ | $0.743\phantom{\rule{3.33333pt}{0ex}}[0.597,\phantom{\rule{3.33333pt}{0ex}}0.889]$ | $0.660\phantom{\rule{3.33333pt}{0ex}}[0.531,\phantom{\rule{3.33333pt}{0ex}}0.790]$ | $0.664\phantom{\rule{3.33333pt}{0ex}}[0.527,\phantom{\rule{3.33333pt}{0ex}}0.802]$ | |

GMM-RF | Dysplastic BE | $0.948\phantom{\rule{3.33333pt}{0ex}}[0.907,\phantom{\rule{3.33333pt}{0ex}}0.989]$ | $0.985\phantom{\rule{3.33333pt}{0ex}}[0.967,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.921\phantom{\rule{3.33333pt}{0ex}}[0.834,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.929\phantom{\rule{3.33333pt}{0ex}}[0.843,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.914\phantom{\rule{3.33333pt}{0ex}}[0.856,\phantom{\rule{3.33333pt}{0ex}}0.972]$ |

Non-dysplastic BE | $0.941\phantom{\rule{3.33333pt}{0ex}}[0.903,\phantom{\rule{3.33333pt}{0ex}}0.979]$ | $0.983\phantom{\rule{3.33333pt}{0ex}}[0.965,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.892\phantom{\rule{3.33333pt}{0ex}}[0.776,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.947\phantom{\rule{3.33333pt}{0ex}}[0.912,\phantom{\rule{3.33333pt}{0ex}}0.982]$ | $0.910\phantom{\rule{3.33333pt}{0ex}}[0.840,\phantom{\rule{3.33333pt}{0ex}}0.981]$ | |

Squamous | $0.993\phantom{\rule{3.33333pt}{0ex}}[0.984,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.999\phantom{\rule{3.33333pt}{0ex}}[0.997,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.985\phantom{\rule{3.33333pt}{0ex}}[0.960,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.988\phantom{\rule{3.33333pt}{0ex}}[0.959,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.986\phantom{\rule{3.33333pt}{0ex}}[0.967,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | |

Weighted Average | $0.952\phantom{\rule{3.33333pt}{0ex}}[0.915,\phantom{\rule{3.33333pt}{0ex}}0.989]$ | $0.986\phantom{\rule{3.33333pt}{0ex}}[0.970,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.955\phantom{\rule{3.33333pt}{0ex}}[0.930,\phantom{\rule{3.33333pt}{0ex}}0.980]$ | $0.941\phantom{\rule{3.33333pt}{0ex}}[0.903,\phantom{\rule{3.33333pt}{0ex}}0.979]$ | $0.942\phantom{\rule{3.33333pt}{0ex}}[0.904,\phantom{\rule{3.33333pt}{0ex}}0.981]$ | |

GMM-SVM | Dysplastic BE | $0.937\phantom{\rule{3.33333pt}{0ex}}[0.913,\phantom{\rule{3.33333pt}{0ex}}0.961]$ | $0.988\phantom{\rule{3.33333pt}{0ex}}[0.980,\phantom{\rule{3.33333pt}{0ex}}0.997]$ | $0.814\phantom{\rule{3.33333pt}{0ex}}[0.708,\phantom{\rule{3.33333pt}{0ex}}0.921]$ | $0.976\phantom{\rule{3.33333pt}{0ex}}[0.948,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.879\phantom{\rule{3.33333pt}{0ex}}[0.811,\phantom{\rule{3.33333pt}{0ex}}0.946]$ |

Non-dysplastic BE | $0.931\phantom{\rule{3.33333pt}{0ex}}[0.909,\phantom{\rule{3.33333pt}{0ex}}0.954]$ | $0.973\phantom{\rule{3.33333pt}{0ex}}[0.959,\phantom{\rule{3.33333pt}{0ex}}0.987]$ | $0.937\phantom{\rule{3.33333pt}{0ex}}[0.882,\phantom{\rule{3.33333pt}{0ex}}0.991]$ | $0.862\phantom{\rule{3.33333pt}{0ex}}[0.814,\phantom{\rule{3.33333pt}{0ex}}0.910]$ | $0.895\phantom{\rule{3.33333pt}{0ex}}[0.858,\phantom{\rule{3.33333pt}{0ex}}0.933]$ | |

Squamous | $0.994\phantom{\rule{3.33333pt}{0ex}}[0.987,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $1.000\phantom{\rule{3.33333pt}{0ex}}[1.000,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $1.000\phantom{\rule{3.33333pt}{0ex}}[1.000,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.959\phantom{\rule{3.33333pt}{0ex}}[0.908,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.978\phantom{\rule{3.33333pt}{0ex}}[0.950,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | |

Weighted Average | $0.950\phantom{\rule{3.33333pt}{0ex}}[0.928,\phantom{\rule{3.33333pt}{0ex}}0.972]$ | $0.986\phantom{\rule{3.33333pt}{0ex}}[0.977,\phantom{\rule{3.33333pt}{0ex}}0.995]$ | $0.942\phantom{\rule{3.33333pt}{0ex}}[0.921,\phantom{\rule{3.33333pt}{0ex}}0.964]$ | $0.931\phantom{\rule{3.33333pt}{0ex}}[0.909,\phantom{\rule{3.33333pt}{0ex}}0.954]$ | $0.933\phantom{\rule{3.33333pt}{0ex}}[0.912,\phantom{\rule{3.33333pt}{0ex}}0.954]$ |

Number of Codeblock | KM | GMM | ||
---|---|---|---|---|

RF | SVM | RF | SVM | |

3 | $0.583\phantom{\rule{3.33333pt}{0ex}}[0.454,\phantom{\rule{3.33333pt}{0ex}}0.713]$ | $0.594\phantom{\rule{3.33333pt}{0ex}}[0.440,\phantom{\rule{3.33333pt}{0ex}}0.748]$ | $0.551\phantom{\rule{3.33333pt}{0ex}}[0.465,\phantom{\rule{3.33333pt}{0ex}}0.637]$ | $0.651\phantom{\rule{3.33333pt}{0ex}}[0.538,\phantom{\rule{3.33333pt}{0ex}}0.764]$ |

5 | $0.690\phantom{\rule{3.33333pt}{0ex}}[0.578,\phantom{\rule{3.33333pt}{0ex}}0.802]$ | $0.698\phantom{\rule{3.33333pt}{0ex}}[0.578,\phantom{\rule{3.33333pt}{0ex}}0.817]$ | $0.681\phantom{\rule{3.33333pt}{0ex}}[0.586,\phantom{\rule{3.33333pt}{0ex}}0.775]$ | $0.698\phantom{\rule{3.33333pt}{0ex}}[0.597,\phantom{\rule{3.33333pt}{0ex}}0.799]$ |

10 | $0.799\phantom{\rule{3.33333pt}{0ex}}[0.685,\phantom{\rule{3.33333pt}{0ex}}0.913]$ | $0.795\phantom{\rule{3.33333pt}{0ex}}[0.682,\phantom{\rule{3.33333pt}{0ex}}0.907]$ | $0.812\phantom{\rule{3.33333pt}{0ex}}[0.716,\phantom{\rule{3.33333pt}{0ex}}0.908]$ | $0.789\phantom{\rule{3.33333pt}{0ex}}[0.685,\phantom{\rule{3.33333pt}{0ex}}0.894]$ |

20 | $0.806\phantom{\rule{3.33333pt}{0ex}}[0.676,\phantom{\rule{3.33333pt}{0ex}}0.936]$ | $0.801\phantom{\rule{3.33333pt}{0ex}}[0.682,\phantom{\rule{3.33333pt}{0ex}}0.920]$ | $0.894\phantom{\rule{3.33333pt}{0ex}}[0.815,\phantom{\rule{3.33333pt}{0ex}}0.973]$ | $0.827\phantom{\rule{3.33333pt}{0ex}}[0.747,\phantom{\rule{3.33333pt}{0ex}}0.906]$ |

50 | $0.821\phantom{\rule{3.33333pt}{0ex}}[0.690,\phantom{\rule{3.33333pt}{0ex}}0.952]$ | $0.793\phantom{\rule{3.33333pt}{0ex}}[0.664,\phantom{\rule{3.33333pt}{0ex}}0.921]$ | $0.931\phantom{\rule{3.33333pt}{0ex}}[0.875,\phantom{\rule{3.33333pt}{0ex}}0.987]$ | $0.843\phantom{\rule{3.33333pt}{0ex}}[0.764,\phantom{\rule{3.33333pt}{0ex}}0.921]$ |

100 | $\mathbf{0}.\mathbf{836}[0.715,\phantom{\rule{3.33333pt}{0ex}}0.957]$ | $0.796\phantom{\rule{3.33333pt}{0ex}}[0.683,\phantom{\rule{3.33333pt}{0ex}}0.909]$ | $0.956\phantom{\rule{3.33333pt}{0ex}}[0.914,\phantom{\rule{3.33333pt}{0ex}}0.997]$ | $0.922\phantom{\rule{3.33333pt}{0ex}}[0.872,\phantom{\rule{3.33333pt}{0ex}}0.972]$ |

150 | $0.824\phantom{\rule{3.33333pt}{0ex}}[0.968,\phantom{\rule{3.33333pt}{0ex}}0.949]$ | $0.809\phantom{\rule{3.33333pt}{0ex}}[0.685,\phantom{\rule{3.33333pt}{0ex}}0.933]$ | $\mathbf{0}.\mathbf{986}[0.970,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $\mathbf{0}.\mathbf{986}[0.977,\phantom{\rule{3.33333pt}{0ex}}0.995]$ |

200 | $0.833\phantom{\rule{3.33333pt}{0ex}}[0.711,\phantom{\rule{3.33333pt}{0ex}}0.957]$ | $\mathbf{0}.\mathbf{812}[0.695,\phantom{\rule{3.33333pt}{0ex}}0.929]$ | $0.984\phantom{\rule{3.33333pt}{0ex}}[0.967,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.978\phantom{\rule{3.33333pt}{0ex}}[0.962,\phantom{\rule{3.33333pt}{0ex}}0.994]$ |

300 | $0.830\phantom{\rule{3.33333pt}{0ex}}[0.701,\phantom{\rule{3.33333pt}{0ex}}0.959]$ | $\mathbf{0}.\mathbf{812}[0.693,\phantom{\rule{3.33333pt}{0ex}}0.931]$ | $0.984\phantom{\rule{3.33333pt}{0ex}}[0.968,\phantom{\rule{3.33333pt}{0ex}}1.000]$ | $0.983\phantom{\rule{3.33333pt}{0ex}}[0.971,\phantom{\rule{3.33333pt}{0ex}}0.995]$ |

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

## Share and Cite

**MDPI and ACS Style**

Sali, R.; Moradinasab, N.; Guleria, S.; Ehsan, L.; Fernandes, P.; Shah, T.U.; Syed, S.; Brown, D.E.
Deep Learning for Whole-Slide Tissue Histopathology Classification: A Comparative Study in the Identification of Dysplastic and Non-Dysplastic Barrett’s Esophagus. *J. Pers. Med.* **2020**, *10*, 141.
https://doi.org/10.3390/jpm10040141

**AMA Style**

Sali R, Moradinasab N, Guleria S, Ehsan L, Fernandes P, Shah TU, Syed S, Brown DE.
Deep Learning for Whole-Slide Tissue Histopathology Classification: A Comparative Study in the Identification of Dysplastic and Non-Dysplastic Barrett’s Esophagus. *Journal of Personalized Medicine*. 2020; 10(4):141.
https://doi.org/10.3390/jpm10040141

**Chicago/Turabian Style**

Sali, Rasoul, Nazanin Moradinasab, Shan Guleria, Lubaina Ehsan, Philip Fernandes, Tilak U. Shah, Sana Syed, and Donald E. Brown.
2020. "Deep Learning for Whole-Slide Tissue Histopathology Classification: A Comparative Study in the Identification of Dysplastic and Non-Dysplastic Barrett’s Esophagus" *Journal of Personalized Medicine* 10, no. 4: 141.
https://doi.org/10.3390/jpm10040141