CortiNet: A Physics-Perception Hybrid Cortical-Inspired Dual-Stream Network for Gallbladder Disease Diagnosis from Ultrasound

TL;DR

CortiNet addresses ultrasound-based gallbladder disease diagnosis by integrating a physics-informed, wavelet-based multi-scale decomposition with a cortex-inspired dual-stream architecture that separately encodes low-frequency structural information and high-frequency perceptual details. A late fusion combines these cues, while an accuracy-guided adaptive inference mechanism selectively activates the perceptual pathway to maintain robustness under speckle noise. The approach yields near-ceiling discriminative performance (class AUCs ≈ 1.0, overall accuracy ≈ 0.9874) with a fraction of the parameters of standard CNNs, and exhibits strong robustness to noise alongside structure-aligned, interpretable Grad-CAM explanations. Together, these features promise reliable, real-time gallbladder disease diagnosis in resource-constrained clinical settings, with potential generalization to other ultrasound tasks.

Abstract

Ultrasound imaging is the primary diagnostic modality for detecting Gallbladder diseases due to its non-invasive nature, affordability, and wide accessibility. However, the low resolution and speckle noise inherent to ultrasound images hinder diagnostic reliability, prompting the use of large convolutional neural networks that are difficult to deploy in routine clinical settings. In this work, we propose CortiNet, a lightweight, cortical-inspired dual-stream neural architecture for gallbladder disease diagnosis that integrates physically interpretable multi-scale signal decomposition with perception-driven feature learning. Inspired by parallel processing pathways in the human visual cortex, CortiNet explicitly separates low-frequency structural information from high-frequency perceptual details and processes them through specialized encoding streams. By operating directly on structured, frequency-selective representations rather than raw pixel intensities, the architecture embeds strong physics-based inductive bias, enabling efficient feature learning with a significantly reduced parameter footprint. A late-stage cortical-style fusion mechanism integrates complementary structural and textural cues while preserving computational efficiency. Additionally, we propose a structure-aware explainability framework wherein gradient-weighted class activation mapping is only applied to the structural branch of the proposed CortiNet architecture. This choice allows the model to only focus on the structural features, making it robust against speckle noise. We evaluate CortiNet on 10,692 expert-annotated images spanning nine clinically relevant gallbladder disease categories. Experimental results demonstrate that CortiNet achieves high diagnostic accuracy (98.74%) with only a fraction of the parameters required by conventional deep convolutional models.

Figures (6)

• Figure 1: Overview of the proposed CortiNet architecture for ultrasound-based gallbladder disease diagnosis. The framework adopts a cortical-inspired dual-stream design that explicitly separates structural and perceptual information pathways. The input ultrasound image is first transformed into structured, frequency-selective representations using multi-scale signal decomposition, yielding complementary low-frequency components that encode global anatomical structure and high-frequency components that capture fine-grained textural and edge information. These components are processed in parallel by two specialized encoding streams with lightweight convolutional blocks, enabling efficient and targeted feature extraction under ultrasound-specific noise and resolution constraints. To integrate complementary cues, a late-stage cortical-style fusion module combines the learned structural and perceptual representations while preserving computational efficiency and minimizing parameter growth. The fused representation is subsequently passed to a compact classification head that produces multi-class gallbladder disease predictions. By operating on multi-resolution representations and embedding strong domain-aware inductive biases, CortiNet achieves robust diagnostic performance with a significantly reduced parameter footprint, making it suitable for deployment in resource-constrained and point-of-care clinical settings.
• Figure 2: Clinical context and dataset overview for gallbladder disease diagnosis. The figure provides an integrated view of the anatomical relevance, imaging modality, and disease diversity represented in the dataset used in this study. Left: Representative ultrasound images illustrating the wide visual heterogeneity across gallbladder disease categories, including benign, inflammatory, and malignant conditions, highlighting challenges posed by speckle noise, low contrast, and subtle morphological differences. Center: Schematic illustration of the human biliary system with the gallbladder highlighted, situating the computational task within its physiological and anatomical context. Right (top): Schematic depiction of the spectrum of gallbladder pathologies considered, encompassing nine distinct disease categories spanning common conditions such as gallstones and cholecystitis to rarer but clinically aggressive entities such as carcinoma. Right (bottom): Summary of dataset composition across disease categories, demonstrating coverage of both high-prevalence and low-prevalence conditions with moderate class imbalance and no single class dominance. Together, these panels emphasize the clinical realism, diagnostic diversity, and representativeness of the ultrasound cohort underpinning the proposed computational framework.
• Figure 3: Diagnostic performance of the proposed model across the gallbladder disease spectrum. (a) Inter-class confusion analysis. Confusion matrix for nine gallbladder disease categories showing strong diagonal dominance, indicating high classification accuracy with minimal inter-class confusion, even among sonographically similar conditions.(b) ROC analysis. One-versus-rest ROC curves for all classes exhibit near-ideal profiles with AUCs approaching unity, confirming excellent discriminative capability across decision thresholds. (c) Class-wise misclassification rate. Misclassification rates derived from the confusion matrix remain low across all categories, with slightly higher errors in morphologically challenging conditions, yet without systematic bias, indicating stable and reliable model behavior. (d) Class-wise quantitative performance metrics. Accuracy, precision, recall, specificity, and NPV for each class demonstrate consistently high performance, with several categories achieving near-perfect or perfect sensitivity and accuracy.
• Figure 4: Comparative diagnostic performance and computational efficiency of CortiNet and baseline models. (a) Overall classification performance across representative convolutional, lightweight, and gallbladder-specific deep learning architectures, evaluated using accuracy, recall, precision, F1-score, and AUC. CortiNet consistently achieves the highest performance across all metrics. (b) Model complexity comparison in terms of trainable parameters, highlighting that CortiNet yields the most accurate result at substantially reduced parameters relative to both general-purpose and task-specific models. (c) Inference time per image, demonstrating that CortiNet delivers superior diagnostic accuracy with the lowest latency, supporting real-time and resource-constrained clinical deployment.
• Figure 5: Ablation and noise robustness analysis of the proposed adaptive dual-branch architecture. (a) Quantitative ablation study across key performance metrics comparing raw pixel input, perceptual (high-frequency) branch only, structural (low-frequency) branch only, and the full CortiNet model. While both branches contribute to performance, their combination yields consistently superior accuracy, F1-score, and AUC, highlighting complementary roles of structural and perceptual representations. (b) Schematic illustration of the proposed noise-aware adaptive architecture, in which a structural branch performs robust low-frequency analysis while a perceptual branch captures fine-grained detail; adaptive suppression attenuates perceptual contributions under high-noise conditions to ensure stable predictions. (c) Sensitivity of classification accuracy to increasing noise strength for CortiNet and representative gallbladder-focused models. CortiNet exhibits the smallest accuracy degradation at higher noise levels, indicating enhanced robustness arising from adaptive branch selection and structure-dominant reasoning.