The Problem Existing deep learning models for HCC and cholangiocarcinoma (CCA) classification suffer from high computational complexity and demanding resource requirements, limiting their deployment in resource-constrained clinical settings where fast, accurate diagnosis is most needed.
The Innovation HTRecNet is a novel CNN architecture built on ResNet50 but modified with two enhancements - SPConv (Split-based Convolution for efficiency) and CBAM (Convolutional Block Attention Module for improved feature selection) - achieving higher accuracy with approximately 30% fewer computations than the baseline.
Performance On a dataset of 5,432 histopathological images (3,380 HCC, 1,536 normal liver, 180 CCA) from Roboflow Universe, HTRecNet achieved 5-fold cross-validation accuracy of 0.993 and an external test set accuracy of 0.97 with Matthews Correlation Coefficient of 0.95 - robust performance across both primary validation approaches.
Three-Class Problem The model simultaneously distinguishes HCC, CCA, and normal liver tissue (Norm-L) - covering the two most common primary liver cancers and providing a built-in normalcy check that would flag cases where neither tumor is detected.
SPConv: Efficiency Through Feature Splitting Split-based Convolution (SPConv) replaces the standard 3x3 convolution in ResNet50 residual blocks by splitting the input feature map into a 'representative' part (processed with a 3x3 kernel capturing key features) and a 'redundant' part (processed with a lightweight 1x1 kernel). A global average pooling attention mechanism fuses the two outputs, reducing computational load without sacrificing discriminative power.
CBAM: Attention to What Matters The Convolutional Block Attention Module (CBAM) applies sequential channel and spatial attention before each residual connection. Channel attention emphasizes which feature channels carry the most diagnostic information (e.g., nuclear morphology channels vs. background staining). Spatial attention then highlights which image regions within those channels are most diagnostically relevant - focusing the model on tumor architecture rather than uninformative areas.
Architecture Integration The final HTRecNet residual block follows the sequence: 1x1 convolution - 3x3 SPConv - 1x1 convolution - CBAM - residual connection. This structure preserves ResNet50's hierarchical feature extraction while adding efficiency (SPConv) and diagnostic focus (CBAM) at each block.
Data Preparation All 5,432 images from the Aperio AT2 scanner at 20x-40x magnification were preprocessed to 224x224 pixels. Data augmentation included elastic transformation (simulating tissue deformation), horizontal and vertical flipping, and random cropping. Gamma correction and median filtering were applied to enhance diagnostic features and reduce noise.
Five-Fold Cross-Validation Results HTRecNet achieved the highest scores across all metrics in five-fold cross-validation: Accuracy 0.993, Precision 0.990, Recall 0.984, F1 0.985, and MCC 0.986 - outperforming the baseline ResNet50 (Accuracy 0.989), ResCBANet (0.991), and ResSPNet (0.986) on every metric.
External Test Set Performance On the 336-image external independent test set (12 CCA, 224 HCC, 100 Norm-L), HTRecNet achieved accuracy of 0.97 and MCC of 0.95, confirming that cross-validation performance was not inflated by overfitting and that the model generalizes to truly unseen data.
Computational Efficiency HTRecNet achieved approximately 30% reduction in both FLOPs (floating point operations) and number of model parameters compared to the baseline ResNet50 - a significant efficiency gain that makes real-time deployment on clinical-grade hardware more feasible without sacrificing performance.
Stability Across Epochs Box plots of performance distributions across all training epochs showed that HTRecNet maintained more consistent performance with less fluctuation than baseline models, suggesting that the attention mechanisms and split convolution improve gradient flow and training stability.
CCA is Commonly Missed CCA constitutes only 10-15% of primary liver cancers, meaning it is vastly underrepresented in pathology practice compared to HCC. Its rarer occurrence makes pathologist expertise with CCA histology less consistent. An AI tool that specifically addresses CCA - even with the limited 180-image CCA training set here - provides a valuable safety net for less-experienced pathologists.
Resource-Limited Deployment The 30% reduction in computational requirements compared to ResNet50 is clinically significant. Many pathology labs in HCC-endemic regions lack high-end GPU workstations. A more computationally efficient model can run on moderate hardware, expanding access to AI-assisted diagnosis in underserved settings.
Treatment Decision Stakes The distinction between HCC and CCA determines eligibility for liver transplantation (approved for HCC within Milan criteria, contraindicated for CCA), choice of resection margin (wider for CCA), and systemic therapy selection (immunotherapy for HCC, FGFR-targeted agents for CCA). Accurate AI-assisted classification directly influences life-changing treatment decisions.
Pathologist Augmentation HTRecNet is positioned as a diagnostic assistant, not a replacement. Heatmap visualization (highlighting which image regions drove classification) allows pathologists to rapidly review and validate AI decisions, maintaining clinical accountability while gaining efficiency.
CCA Class Imbalance Only 180 CCA training images and 12 CCA external test images were available. CCA is far rarer than HCC, and the severe class imbalance limits the model's ability to learn the full spectrum of CCA histological patterns. Larger CCA datasets from hepatobiliary surgery centers are needed.
Single Dataset Source The training data came from Roboflow Universe, a single publicly available dataset. This source may not capture the full diversity of staining protocols, scanner hardware, and population-specific tumor morphology encountered across global pathology practices.
Three-Class Scope The model distinguishes HCC, CCA, and normal liver but does not address metastatic liver tumors - the most common hepatic malignancy in many Western countries. Expanding to include mCRC, breast, and lung metastases would substantially increase clinical utility.
Prospective Multi-Center Validation Clinical deployment requires validation on prospective multi-institutional datasets with histopathologically confirmed diagnoses. Future studies should also assess HTRecNet performance on biopsy specimens (as opposed to resection specimens), which present smaller tissue samples with more limited histological context.