Deep learning in differentiating the colorectal cancer combined with hepatic enhancing nodules: liver metastases vs hemangiomas

TL;DR

A deep learning model helps radiologists distinguish between liver metastases and hemangiomas in colorectal cancer patients using CT scans.

Contribution

The study introduces a deep learning model that improves diagnostic accuracy for liver lesions in CRC patients.

Findings

The ResNet-152 model achieved an AUC of 0.875 for classifying CRLMs and HMs.

DL assistance improved radiologists' performance for 10–30 mm lesions but not for subcentimeter lesions.

TotalSegmentator showed lower segmentation consistency for subcentimeter lesions.

Abstract

To assess a deep learning (DL) model using portal-venous phase CT for discriminating colorectal cancer liver metastasis (CRLMs) and hemangiomas (HMs).

Colorectal cancer (CRC) patients diagnosed with CRLMs or HMs at two medical centers from January 2018 and April 2024 were retrospectively included. Lesions were automatically segmented using TotalSegmentator. DL models, DenseNet-201 and ResNet-152, were trained to classify CRLMs and HMs. Their performance, measured by AUC, was evaluated on validation and test sets. Subgroup analyses were conducted for lesions ≤ 10 mm (subcentimeter) and 10–30 mm. Radiologists’ diagnostic performance with and without DL assistance was compared using a multi-reader multi-case analysis.

534 CRLMs (134 CRC-patients; median, 60 years) and 262 HMs (154 CRC-patients; median, 62 years) were divided into the training, validation and test set. The Dice coefficients of TotalSegmentor for automatically segmenting subcentimeter and 10–30 mm lesions were 0.692 ± 0.099 and 0.861 ± 0.033, respectively (p < 0.01). ResNet-152 model achieved AUCs of 0.875 (95% CI: 0.838–0.912), 0.858 (95% CI: 0.781–0.935), 0.776 (95% CI: 0.703–0.848) for classifying CRLMs and HMs on the training, validation, and test sets, respectively. The AUCs for distinguishing between 10–30 mm CRLMs and HMs improved from 0.851 (95% CI: 0.821–0.880) to 0.879 (95% CI: 0.853–0.906) with DL assistance compared to without (p = 0.015). For subcentimeter CRLMs and HMs, the AUCs for the radiologists and the DL-assisted diagnosis were 0.742 (95% CI: 0.669–0.814) and 0.763 (95% CI: 0.681–0.845), respectively (p = 0.558).

DL can assist radiologists in distinguishing 10–30 mm CRLMs from HMs in CRC patients. The value of DL-assisted diagnosis is limited for subcentimetre CRLMs and HMs.

Dynamic detection of hypoenhancing liver lesions in patients with CRC is exceptionally challenging. The DL tool we have developed can assist in evaluating CRLMs and HMs.

TotalSegmentator can perform automatic segmentation of CRLMs and HMs, but demonstrates poorer segmentation consistency for subcentimeter lesions.This DL model assists radiologists in distinguishing 10–30 mm CRLMs from HMs in CRC patients.Subcentimeter CRLMs and HMs can require further MRI scanning.

TotalSegmentator can perform automatic segmentation of CRLMs and HMs, but demonstrates poorer segmentation consistency for subcentimeter lesions.

This DL model assists radiologists in distinguishing 10–30 mm CRLMs from HMs in CRC patients.

Subcentimeter CRLMs and HMs can require further MRI scanning.

Figures