Precise Liver Tumor Segmentation in CT Using a Hybrid Deep Learning-Radiomics Framework

arXiv ID: 2512.07574v1

Published: 2025-12-08

Authors: Xuecheng Li, Weikuan Jia, Komildzhon Sharipov, Alimov Ruslan, Lutfuloev Mazbutdzhon, Ismoilov Shuhratjon, Yuanjie Zheng

Categories: eess.IV, cs.CR, cs.CV

Relevance Score: 0.98 / 1.00

View on arXiv Download PDF

Summary

This paper introduces a hybrid deep learning-radiomics framework designed for precise, joint liver and liver-tumor segmentation in contrast-enhanced CT images. The proposed multi-stage approach combines an attention-enhanced cascaded U-Net for initial probability maps, inter-slice consistency enforcement, radiomics-based false-positive rejection, and a 3D patch-based CNN for final boundary refinement. This aims to overcome challenges of manual contouring, such as slowness and observer dependency, and common auto-segmentation difficulties like low contrast and blurred boundaries.

Medical Relevance

Accurate three-dimensional delineation of liver tumors on CT is fundamental for personalized treatment planning, precise surgical or interventional navigation, and reliable assessment of treatment response, all of which directly impact patient outcomes and therapeutic strategies. This work aims to provide a standardized and efficient method for this crucial task.

AI Health Application

The AI application is the automated, precise three-dimensional segmentation of liver tumors and the liver parenchyma from CT scans. This technology aims to overcome limitations of manual contouring (slowness, observer dependency, lack of standardization) by using deep learning (U-Net, CNNs) and radiomics. Its direct impact is to improve accuracy and efficiency in cancer diagnosis, treatment planning (e.g., radiation therapy, surgery), intraoperative navigation, and objective assessment of treatment response for liver cancer patients.

Key Points

A novel hybrid framework integrates deep learning (attention-enhanced cascaded U-Net, 3D CNN) and handcrafted radiomics for precise, joint liver and liver-tumor segmentation on CT scans.
Initial liver and tumor probability maps are generated by a 2.5D two-stage network featuring a densely connected encoder, sub-pixel convolution decoders, and multi-scale attention gates.
Inter-slice temporal consistency is enforced through a simple three-slice refinement rule along the cranio-caudal direction, which effectively restores thin/tiny lesions and suppresses isolated noise.
False-positive regions are robustly rejected by a Random Forest classifier, trained on 20 stable and highly informative radiomic features selected from an initial set of 728 descriptors (spanning intensity, texture, shape, boundary, and wavelet groups).
Final voxel-level relabeling and contour smoothing around the tumor boundary are achieved by a compact 3D patch-based CNN, derived from AlexNet, operating in a narrow band.
The framework is designed to specifically address inherent difficulties in automatic liver tumor segmentation, including low lesion-parenchyma contrast, blurred boundaries, heterogeneous enhancement patterns, and confounding structures.

Methodology

The proposed method is a sequential multi-stage hybrid framework. It begins with a 2.5D attention-enhanced cascaded U-Net that takes short stacks of axial slices as input to produce initial liver and tumor probability maps. This is followed by a three-slice refinement rule applied along the cranio-caudal axis to enforce inter-slice consistency, restoring small lesions and suppressing noise. Next, a radiomics module extracts 728 descriptors (intensity, texture, shape, boundary, wavelet) from candidate lesions, reduces them to 20 stable, highly informative features via multi-strategy selection, and uses a Random Forest classifier to reject false positives. Finally, a compact 3D patch-based CNN (derived from AlexNet) performs voxel-level relabeling and contour smoothing in a narrow band around the tumor boundary.

Key Findings

The hybrid framework effectively addresses the multi-faceted challenges of liver tumor segmentation by: 1) providing robust initial probability maps using an attention-enhanced U-Net, 2) improving lesion detection and suppressing noise through inter-slice temporal consistency, 3) accurately rejecting false-positive regions via a radiomics-driven Random Forest classifier, and 4) achieving precise and smooth tumor boundaries with a dedicated 3D patch-based CNN.

Clinical Impact

This automated and standardized segmentation method has the potential to significantly enhance the efficiency and consistency of liver tumor delineation in clinical practice, thereby reducing clinician workload and inter-observer variability. This directly facilitates more accurate and personalized treatment planning for radiation therapy and surgery, improves guidance for targeted ablation procedures, and enables objective, quantifiable assessment of therapy efficacy, ultimately leading to improved patient care and clinical decision-making.

Limitations

The abstract does not explicitly state limitations regarding dataset size, computational performance, generalizability across different CT scanner manufacturers or protocols, or detailed quantitative performance metrics for the proposed framework.

Future Directions

The abstract does not explicitly suggest future research directions for the framework.

Medical Domains

Diagnostic Radiology Oncology Surgical Planning Radiation Oncology Interventional Radiology

Keywords

Liver tumor segmentation CT imaging Deep learning Radiomics U-Net 3D CNN Medical image analysis

Abstract

Accurate three-dimensional delineation of liver tumors on contrast-enhanced CT is a prerequisite for treatment planning, navigation and response assessment, yet manual contouring is slow, observer-dependent and difficult to standardise across centres. Automatic segmentation is complicated by low lesion-parenchyma contrast, blurred or incomplete boundaries, heterogeneous enhancement patterns, and confounding structures such as vessels and adjacent organs. We propose a hybrid framework that couples an attention-enhanced cascaded U-Net with handcrafted radiomics and voxel-wise 3D CNN refinement for joint liver and liver-tumor segmentation. First, a 2.5D two-stage network with a densely connected encoder, sub-pixel convolution decoders and multi-scale attention gates produces initial liver and tumor probability maps from short stacks of axial slices. Inter-slice temporal consistency is then enforced by a simple three-slice refinement rule along the cranio-caudal direction, which restores thin and tiny lesions while suppressing isolated noise. Next, 728 radiomic descriptors spanning intensity, texture, shape, boundary and wavelet feature groups are extracted from candidate lesions and reduced to 20 stable, highly informative features via multi-strategy feature selection; a random forest classifier uses these features to reject false-positive regions. Finally, a compact 3D patch-based CNN derived from AlexNet operates in a narrow band around the tumor boundary to perform voxel-level relabelling and contour smoothing.