A CT-based deep learning for segmenting tumors and predicting microsatellite instability in patients with colorectal cancers: a multicenter cohort study

Weicui Chen ¹, Kaiyi Zheng ^2,3, Wenjing Yuan ¹

¹Radiology Department, The Second Affiliated Hospital of Guangzhou University of Chinese Medicine, Guangzhou, 510120, Guangdong, China.
²School of Biomedical Engineering, Southern Medical University, Guangzhou, 510515, Guangdong, China.
³Guangdong Provincial Key Laboratory of Medical Image Processing, Southern Medical University, Guangzhou, 510515, Guangdong, China.
⁴Department of Medical Imaging, Nanfang Hospital, Southern Medical University, Guangzhou, 510515, Guangdong, China.
⁵Radiology Department, Sun Yat-Sen Memorial Hospital, Sun Yat-Sen University, Guangzhou, 510120, Guangdong, China.
⁶Radiology Department, Zhujiang Hospital, Southern Medical University, Guangzhou, 510282, Guangdong, China. zhibowen@163.com.
⁷School of Biomedical Engineering, Southern Medical University, Guangzhou, 510515, Guangdong, China. limingzhongmindy@gmail.com.
⁸Guangdong Provincial Key Laboratory of Medical Image Processing, Southern Medical University, Guangzhou, 510515, Guangdong, China. limingzhongmindy@gmail.com.
⁹Radiology Department, The Second Affiliated Hospital of Guangzhou University of Chinese Medicine, Guangzhou, 510120, Guangdong, China. liuxian74@hotmail.com.

⁰Radiology Department, The Second Affiliated Hospital of Guangzhou University of Chinese Medicine, Guangzhou, 510120, Guangdong, China.

La Radiologia Medica +

|

November 25, 2024

View abstract on PubMed

Summary

Deep learning models effectively segment colorectal cancer (CRC) tumors on CT scans. Combined models integrating imaging and clinical data accurately predict microsatellite instability (MSI) in CRC patients.

Area Of Science

Radiology
Oncology
Artificial Intelligence

Background

Colorectal cancer (CRC) diagnosis and treatment planning benefit from accurate tumor segmentation and microsatellite instability (MSI) prediction.
Deep learning (DL) offers potential for automating these tasks using medical imaging.

Purpose Of The Study

To develop and validate DL models for automated tumor segmentation and MSI prediction in CRC using preoperative contrast-enhanced CT images.
To assess the performance of DL models in comparison to traditional methods.

Main Methods

Retrospective analysis of 2180 CRC patients' CT scans.
Development of an nnU-Net model for tumor auto-segmentation.
Training and validation of ViT or CNN models for MSI prediction using imaging data and/or clinical-pathological factors.
Evaluation using metrics like Dice coefficient, AUC, and decision curve analysis.

Main Results

The segmentation model achieved high performance in the external test set (e.g., Dice coefficient of 0.71).
Combined DL models integrating CT images and clinical data significantly outperformed clinical models and image-only models in MSI prediction (AUCs of 0.83 and 0.82).
Decision curve analysis demonstrated superior clinical utility of the combined models.

Conclusions

Deep learning models significantly improve tumor segmentation efficiency in CRC.
Integrated DL models combining contrast-enhanced CT imaging and clinicopathological data show strong diagnostic performance for MSI prediction in CRC.

Keywords:

Automatic segmentation Colorectal cancer Computed tomography Deep learning Microsatellite instability