BactSegNet: A Growth-Stage-Aware, Searched Instance-Segmentation Model for Salmonella and Escherichia coli Microcolonies in Pure-Culture Bright-Field Microscopy with a Food-Matrix Surrogate

Instance segmentation of small, dense, low-contrast microcolonies against confounding food matrices limits early microscopy-based screening for Salmonella and E. coli. We present BactSegNet, combining EfficientNetV2, BiFPN, selective-kernel attention, and growth-stage FiLM conditioning based on a learnable embedding of recorded acquisition stage. BE-NAS explored a 2304-configuration search space and evaluated 71 unique architectures. Across twelve retrained candidates, proxy and full-training rankings agreed (Spearman $\rho =0.867$ for mAP@50 and $\rho =0.790$ for mAP@50–95). Across the 71 architectures evaluated in the BE-NAS proxy search, three-seed proxy fitness showed high reliability ($\mathrm{ICC}(1,1)=0.91$; 213 seed-level evaluations), while phenotype caching reduced 180 scheduled evaluations to 71 unique trainings (213 rather than 540 seed-level proxy runs). On $14,318$ expert-verified polygon instances spanning six stages, two species, and onion-present/absent conditions, BactSegNet-L achieved mask mAP@50 = $0.9636$ and mAP@50–95 = $0.7154$, $3.56$ percentage points above a multi-seed YOLOv8 baseline and higher mAP@50 than sixteen class-aware single-run references across three paradigms. A capacity-preserving stage-perturbation control reduced mAP@50 from $0.9636\pm 0.0049$ with correct stage assignments to $0.9487\pm 0.0040$ after shuffling image–stage correspondence across three seeds. Gains reached $8.66$ pp for Salmonella at $1.5$ h and $4.90$ pp on onion. Across 2756 matched instances, area MAPE was $12.15\%$ with $r=0.997$, dominated by the largest instances. BactSegNet-Efficient retained $0.9510$ mAP@50 with $19.1\%$ fewer FLOPs at $43.9$ fps. The main dataset comprises pure-culture frames; an independent, expert-annotated mixed-culture evaluation (400 images, 3.5 and 4.0 h) yielded mask mAP@50=$0.8857\pm 0.0206$ across three seeds without retraining. Recorded growth-stage metadata are required at inference, unlike the image-only comparison models. These results support early, physically calibrated microcolony segmentation while distinguishing the original pure-culture evaluation from the additional mixed-culture test.