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

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.

💡 Research Summary

CortiNet is a lightweight, physics‑perception hybrid neural network designed for automated diagnosis of gallbladder diseases from ultrasound (US) images. The authors begin by highlighting the clinical importance of US for gallbladder assessment and the inherent challenges posed by low resolution, speckle noise, and operator dependence. Conventional deep convolutional networks, while powerful, are often over‑parameterized and difficult to deploy in resource‑constrained point‑of‑care settings.

To address these issues, CortiNet integrates a physics‑driven multi‑scale signal decomposition with a biologically inspired dual‑stream architecture. An input US image is first processed by a Daubechies wavelet transform that separates the image into a low‑frequency approximation component (A₁) capturing global anatomical structure, and three high‑frequency detail components (DHL₁, DLH₁, DHH₁) that encode orientation‑specific edge and texture information. This explicit frequency decomposition embeds a strong inductive bias derived from the physics of ultrasound wave‑tissue interaction, making the network more robust to speckle noise.

The architecture then branches into two independent encoders. The Structural Stream (S) receives A₁ and learns coarse shape, organ‑level morphology, and spatial continuity using a shallow convolutional block. The Detail Stream (T) processes the concatenated high‑frequency components and learns fine‑grained texture, edge orientation, and subtle pathological cues through a deeper block. Because the streams have disjoint parameters, high‑frequency noise does not contaminate the structural representation.

After feature extraction, each stream undergoes Global Average Pooling, producing scale‑invariant descriptors f_S and f_T. These vectors are concatenated and fed into a compact fully‑connected classification head with softmax output, yielding probabilities over nine clinically relevant gallbladder disease categories (e.g., cholelithiasis, acute/chronic cholecystitis, polyps, carcinoma). The entire model contains roughly 1.2 million parameters—an order of magnitude fewer than standard ResNet‑50 or Inception‑based models—yet achieves comparable or superior performance.

A notable contribution is the Noise‑Aware Adaptive Inference mechanism. During a brief calibration phase, the structural pathway’s empirical accuracy is measured on a small validation set. If this accuracy falls below a predefined threshold, the detail pathway is deactivated for subsequent inference, reducing computational load without sacrificing robustness. This dynamic gating reflects the clinical reality that high‑frequency details are often unreliable in heavily speckled scans.

For explainability, the authors apply Grad‑CAM exclusively to the structural stream. By avoiding the noisy detail stream, the resulting heatmaps highlight anatomically meaningful regions such as gallbladder wall contours, stone locations, and lesion boundaries, thereby increasing clinician trust.

Experimental evaluation uses a curated dataset of 10,692 expert‑annotated US images collected at the Indian Institute of Technology Delhi, covering nine disease classes. Five‑fold cross‑validation reports an overall accuracy of 98.74 % and an average F1‑score of 0.98. Inference time averages 14 ms per image on a modest GPU, satisfying real‑time requirements. Ablation studies demonstrate that (1) removing the wavelet decomposition drops accuracy by 4.4 %, (2) merging the two streams into a single encoder reduces performance by 2.7 %, and (3) disabling adaptive inference increases FLOPs by ~30 % with negligible accuracy change. These results confirm that each design element contributes meaningfully to the final system.

The paper acknowledges limitations: the fixed wavelet basis may not generalize across all US devices or acquisition settings, and the model has not been tested on rare or atypical gallbladder pathologies beyond the nine categories. Future work is proposed to incorporate learnable multi‑scale filters, transformer‑based global context modules, and multi‑center domain adaptation to broaden applicability.

In summary, CortiNet demonstrates that embedding domain‑specific physics and human‑inspired perceptual processing into a compact dual‑stream network can deliver state‑of‑the‑art diagnostic accuracy, computational efficiency, and clinically relevant explainability for gallbladder disease detection from ultrasound images.