Automated Liver Lesion detection and classification using Deep learning and CT image

Abstract

Automated liver lesion analysis from computed tomography (CT) images can assist clinicians by providing objective information about lesion location and disease category. This paper presents a segmentation-based liver lesion classification framework developed using the Liver Tumor Segmentation Challenge 2017 (LiTS17) dataset. The proposed workflow follows a sequential strategy in which abdominal CT images are first preprocessed, the liver is segmented from the abdominal image, tumors are segmented from the extracted liver region, features are derived from the segmented tumor region, and the lesion is finally classified as benign or malignant. Several CT preprocessing methods are considered, including Hounsfield Unit (HU) windowing, resizing, Contrast Limited Adaptive Histogram Equalization (CLAHE), Min-Max normalization, Z-score normalization and median filtering, For the segmentation stage, multiple U-Net, LinkNet, PSPNet, FPN, DANet, CFNet, SpatialOCRNet and DeepLab configurations are investigated. The reported segmentation experiments achieve their strongest result with CFNet using a ResNet50V2 backbone, with 99.52% Dice and 99.05% IoU. For lesion classification, models are evaluated from scratch and with pretrained weights. Among the reported pretrained models, InceptionResNetV2 achieves the highest classification accuracy of 99%, with benign and malignant F1-scores of 98% and 99%, respectively. The study demonstrates the potential of combining explicit anatomical/lesion segmentation with deep feature learning for liver lesion classification.