Application of Foundation Models for Colorectal Cancer Tissue Classification in Mass Spectrometry Imaging

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This research provides a strong contribution to computational pathology by adapting foundation models. However, there is a limitations in generalizability and clinical integration.  

1. Large multi center data could validate the scalability.
2. Computational cost is missing which plays a significant role in real time surgical applications.
3. Explore spatial aware architectures.
4. Publish code to help researcher to further dive deep in this area.

Author Response

Please see the file attached

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

pls check the comments in the file

Comments for author File: Comments.pdf

Author Response

Please see the file attached

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This study exploreS the application of foundation models, particularly those from the audio-language domain, for the classification of colorectal cancer tissue using Mass Spectrometry Imaging (MSI). The authors have compared multiple models and conducted ablation studies to validate the robustness of their approach under varying data conditions. The results indicate that the FACT model, after specific pre-training, demonstrates excellent performance across all metrics. Notably, its performance remains strong, and even improves, under reduced data conditions.

However, before the manuscript can be accepted for publication, I believe there are several major and minor issues that the authors need to address:

- A very interesting finding in this study is that the FACT and DreaMS models, when fine-tuned on the "small dataset," outperform their performance on the "large dataset". The authors attribute this to the model's inductive bias and generalization capabilities. The authors should further explore the potential reasons for this phenomenon. The "small dataset" was created by "manually reducing the ROI selection". Could this introduce an "expert selection bias"? Specifically, the manually selected regions might contain higher-quality, more representative "essential" spectra, effectively filtering out some noisy or ambiguous samples. If this is the case, the significance of this finding shifts to "high-quality fine-tuning data is crucial" rather than simply "the model performs well on small data." It is recommended that the authors add an analysis of this possibility to their discussion.

All data in this study were generated from a single DESI-MSI system. Mass spectrometry data are highly sensitive to instrument parameters, calibration, and sample preparation. The authors should mention in the limitations section that the model's performance on data from different instruments, laboratories, or under different data acquisition protocols remains unknown. This is crucial for assessing whether the technology can be generalized to diverse clinical settings.

The paper does not mention the time and computational resources required to train these foundation models. Adding information about the hardware configuration (e.g., GPU model) and approximate training times in the "Implementation Details" section of the methods would provide a useful reference for other researchers wishing to reproduce or extend this work.

To enhance the reproducibility of this work for other researchers, the authors should, if possible, make the code and relevant data open-source.

Author Response

Please see the file attached

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

authors improved the manuscript significantly, no more comments.

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have addressed my concerns.