Quantization-Aware Pruned MobileNetV2 for Real-Time Liver Tumor Classification on CT Images
Early diagnosis of liver cancer is critical due to its high mortality rate and late clinical presentation. Computed tomography (CT) remains the standard modality for liver lesion assessment; however, interpretation requires experts that are skilled and prone to diagnostic variability. Deep learning has enhanced liver lesion classification accuracy, though large models come with computational burdens that hinder deployment in clinical and resource-limited setting. This study introduces an optimized MobileNetV2 architecture for liver tumor classification, integrating unstructured pruning and quantization-aware training (QAT) to reduce the model size and latency while maintaining diagnostic performance. Using LiTS17 CT dataset, 17,506 slices were extracted and labeled as tumor or non-tumor. A Baseline finetuned MobileNetV2 achieved 99.6% accuracy but required 26.10 MB and 8 FPS processing. After pruning and weight stripping, size reduced to 8.91 MB with a 31.57 FPS inference rate and 94.8% accuracy. Further QAT yielded an 8-bit TensorFlow Lite model with 8.46 MB size and 94.5% accuracy. These results demonstrate that lightweight, compressed models can support real-time liver tumor diagnosis.
Download Paper