Su Jin Im

Although deep learning methods have shown high performance in chest X-ray classification, high accuracy alone does not guarantee reliable reasoning. A model may still exhibit pathological behavior, such as unstable evidence usage under harmless input changes, inconsistent reasoning across augmented views, excessive dependence on surrounding frame information, and appearance bias dominance, where prediction relies too heavily on intensity while neglecting texture and shape. In this paper, we propose a bio-inspired pathology-aware, factor-aware framework for explainable and reliable chest X-ray classification, inspired by biological vision principles such as figure–ground separation, selective attention, and balanced use of complementary visual cues. During training, the method regularizes appearance bias dominance through evidence-guided counterfactual perturbations that mimic cue-suppression analysis in biological perception, thereby revealing and penalizing excessive factor dependence. During testing, it evaluates model behavior using four criteria: reasoning stability, augmentation inconsistency, appearance bias dominance, and framing sensitivity. This combination enables the framework to go beyond conventional inference-time explanation by both correcting pathological behavior during training and exposing it during evaluation. From a biomimetic perspective, the framework encourages the model to separate relevant foreground anatomy from surrounding background and to avoid over-reliance on a single dominant cue. The proposed approach improves interpretability and reliability without modifying the backbone architecture or increasing model size or inference-time cost. The proposed training process improved the F1-score of DenseNet-121 from 0.899 to 0.931, while also producing more stable and balanced reasoning.

Background/Objectives: Colorectal cancer (CRC) is one of the leading causes of cancer-related mortality worldwide, and accurate histopathological tissue segmentation is critical for timely and reliable diagnosis. Healthcare systems represent complex adaptive environments where diagnostic tools must function reliably across heterogeneous clinical settings, varying staining protocols, and resource-constrained infrastructures. However, existing deep learning segmentation models often require substantial computational resources, limiting their deployment in such settings. This study proposes a novel, resource-efficient colorectal histopathology segmentation network (RCHS-Net) designed for robust clinical deployment across diverse and resource-constrained healthcare environments. Methods: RCHS-Net employs a compact multi-scale encoder with channel recalibration blocks, a gland context module (GCM) with three parallel atrous convolutions and lightweight self-attention for multi-scale contextual feature extraction, and a feature pyramid decoder (FPD) for fine-grained spatial reconstruction. To address the demands of real-world healthcare systems, feature-wise linear modulation (FiLM) conditioning enables class-aware segmentation across multiple tissue categories, while MixStyle augmentation improves stain domain generalization across heterogeneous laboratory and scanner conditions. Results: The model was evaluated on two publicly available benchmark datasets: the EBHI-Seg dataset and the GlaS dataset. On EBHI-Seg, RCHS-Net achieved a mean Dice coefficient of 95.20% and a mean IoU of 91.10% across six colorectal tissue classes, with only 243,226 trainable parameters. On the GlaS benchmark, RCHS-Net attained a Dice score of 93.39% and an IoU of 88.32%, outperforming state-of-the-art methods. Conclusions: RCHS-Net demonstrates that high-accuracy histopathology segmentation can be achieved with a compact architecture, offering a scalable and practical solution for AI-assisted cancer diagnosis across the complex, heterogeneous conditions of real-world healthcare systems, supporting scalable and equitable cancer diagnostics globally.

The deep classifiers used for chest X-ray diagnosis can be sensitive to the visual frame around the evidence, producing correct labels for the wrong reasons by relying on the background, borders, text marks, or acquisition artifacts. This framing sensitivity reduces their trustworthiness, and they can fail under a distribution shift. Inspired by biological vision, especially figure–ground segregation and selective attention, we propose a bio-inspired adversarial attention alignment training process that encourages evidence-centered decisions without changing the classifier structure. A classifier is first trained with image-level labels. Class activation mapping (CAM) is then used to produce a differentiable heatmap that indicates where the model attends. We treat this heatmap as a generated localization map and train a discriminator to distinguish the generated heatmaps from ground truth masks. The classifier is updated using a joint objective that preserves the classification performance while pushing its CAM heatmap toward a mask-like structure, reducing the reliance on background cues. We also introduce evaluation measures for the test phase, including augmentation inconsistency (prediction flip rate under angle-based augmentations) and framing sensitivity (CAM energy outside the mask). The experiments show improved lung-focused attention and robustness, while requiring masks only during training and no additional inputs at inference.