Automated Liver Lesion Detection and Classification on LiTS'17

Abstract

Automated detection and classification of liver lesions from computed tomography (CT) can give radiologists objective, reproducible information about lesion location and malignancy. This paper presents a five-phase framework for benign versus malignant liver lesion classification on the Liver Tumor Segmentation Challenge 2017 (LiTS'17) dataset and evaluates it through two routes: a machine learning (ML) route built on classical segmentation, handcrafted features and conventional classifiers, and a deep learning (DL) route built on advanced segmentation networks and convolutional neural network (CNN) classifiers. Both routes share a pre-processing phase of Hounsfield Unit (HU) windowing, resizing, normalisation, Contrast Limited Adaptive Histogram Equalization (CLAHE), and data augmentation. In the ML route, four segmentation pipelines (marker-controlled watershed with Gaussian Mixture Models, Fuzzy C-Means with Grey Wolf Optimization, FCM with region growing, and Kirsch edge detection with mean-shift smoothing and FCM) are combined with shape, intensity and texture features (FOS, GLCM, GLRLM, LBP, Gabor, SFTA), PCA-based reduction and an ensemble of KNN, MLP, SVM and Random Forest; the best pipeline reaches 96.97% accuracy, 96.5% precision and 96% recall. In the DL route, four advanced segmentation models (DANet, CFNet, SpatialOCRNet and DeepLab) are benchmarked with four backbones; CFNet and DANet give the most consistent masks, with about 99.5% Dice and 99.0% IoU. For classification, ImageNet-pretrained InceptionResNetV2 achieves 99% accuracy with benign and malignant F1-scores of 98% and 99%. The results show that the DL route outperforms the best ML route by about 2 percentage points in accuracy and removes the need for handcrafted feature design, while the ML route remains a competitive and interpretable baseline.