Development and external evaluation of a self-learning auto-segmentation model for Colorectal Cancer Liver Metastases Assessment (COALA)

Abstract

Objectives: Total tumor volume (TTV) is associated with overall and recurrence-free survival in patients with colorectal cancer liver metastases (CRLM). However, the labor-intensive nature of such manual assessments has hampered the clinical adoption of TTV as an imaging biomarker. This study aimed to develop and externally evaluate a CRLM auto-segmentation model on CT scans, to facilitate the clinical adoption of TTV.

Methods: We developed an auto-segmentation model to segment CRLM using 783 contrast-enhanced portal venous phase CTs (CT-PVP) of 373 patients. We used a self-learning setup whereby we first trained a teacher model on 99 manually segmented CT-PVPs from three radiologists. The teacher model was then used to segment CRLM in the remaining 663 CT-PVPs for training the student model. We used the DICE score and the intraclass correlation coefficient (ICC) to compare the student model's segmentations and the TTV obtained from these segmentations to those obtained from the merged segmentations. We evaluated the student model in an external test set of 50 CT-PVPs from 35 patients from the Oslo University Hospital and an internal test set of 21 CT-PVPs from 10 patients from the Amsterdam University Medical Centers.

Results: The model reached a mean DICE score of 0.85 (IQR: 0.05) and 0.83 (IQR: 0.10) on the internal and external test sets, respectively. The ICC between the segmented volumes from the student model and from the merged segmentations was 0.97 on both test sets.

Conclusion: The developed colorectal cancer liver metastases auto-segmentation model achieved a high DICE score and near-perfect agreement for assessing TTV.

Critical relevance statement: AI model segments colorectal liver metastases on CT with high performance on two test sets. Accurate segmentation of colorectal liver metastases could facilitate the clinical adoption of total tumor volume as an imaging biomarker for prognosis and treatment response monitoring.

Key points: Developed colorectal liver metastases segmentation model to facilitate total tumor volume assessment. Model achieved high performance on internal and external test sets. Model can improve prognostic stratification and treatment planning for colorectal liver metastases.

Keywords: Artificial intelligence; Biomarkers; Colorectal neoplasms; Liver; Tumor.

Fig. 1

Example of a manually segmented portal venous phase axial computed tomography scan performed by a trio of radiologists and combined using the STAPLE algorithm. Green = CRLM, turquoise = liver, pink = spleen, dark blue = pancreas, light blue = adrenal glands, red = stomach, yellow = colon

Fig. 2

The proposed learning framework for CRLM and abdominal organ segmentation on contrast-enhanced CT scans. CT-PVP, portal venous phase computed tomography scan

Fig. 3

Comparison between the COALA model’s segmentation and the merged segmentation within a portal venous phase axial computed tomography scan. Red = automatic segmentation performed by our model, green = merged manual segmentation performed by three radiologists and merged using the STAPLE algorithm