An Explainable Deep Learning Framework with Multi-Scale Backbones for 9-Class Gallbladder Pathology Identification in Ultrasound Scans

Autonomous Full-Frame Pipeline: Evaluates multi-class gallbladder disease identification directly on uncropped, full-frame ultrasound scans, removing the manual seed-point bottleneck of Active Contour Segmentation (ACS).
Luminance-Preserving Contrast Enhancement: Applies CLAHE strictly to the luminance ($L^*$) channel in the CIE $L^*a^*b^*$ color space, enhancing mucosal margins and posterior shadowing without introducing chrominance distortion.
Comprehensive Architecture Benchmark: Evaluates ResNet-50, EfficientNet-B0, and MobileNetV3-Large across a 13,872-scan cohort using inverse-frequency class-weighted cross-entropy loss and a two-stage transfer learning protocol. High-Stakes Diagnostic Reliability: Achieves a 100% true positive rate (1.00 recall, 1.00 precision, and 1.000 ROC-AUC) across six acute conditions (gallstones, normal scans, acute cholecystitis, gangrenous cholecystitis, perforation, and polyps). Radiological Dilemma Mapping: Characterizes diagnostic ambiguity as strictly confined to the adenomyomatosis–carcinoma–wall thickening triad, directly mirroring clinical challenges in abdominal radiology. Ultra-Fast Edge Viability: Demonstrates real-time POCUS bedside triage potential via MobileNetV3-Large, yielding an 83.60% macro F1-score with only 3.47M parameters and a 6.00 ms GPU latency ($>$160 frames per second). Transparent Saliency Verification: Validates decision pathways using Grad-CAM heatmaps, confirming network reliance on genuine sonographic biomarkers rather than machine telemetry or peripheral calipers.