Distinguishing pure histopathological growth patterns of colorectal liver metastases on CT using deep learning and radiomics: a pilot study

• Published in: Clinical & experimental metastasis Vol. 38; no. 5; pp. 483 - 494
• Main Authors: Starmans, Martijn P. A., Buisman, Florian E., Renckens, Michel, Willemssen, François E. J. A., van der Voort, Sebastian R., Groot Koerkamp, Bas, Grünhagen, Dirk J., Niessen, Wiro J., Vermeulen, Peter B., Verhoef, Cornelis, Visser, Jacob J., Klein, Stefan
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 01.10.2021; Springer Nature B.V
• Subjects: Aged; Artificial neural networks; Biomarkers; Biomedical and Life Sciences; Biomedicine; Cancer Research; Colorectal Neoplasms - pathology; Computed tomography; Correlation coefficients; Deep Learning; Evaluation; Female; Growth patterns; Hematology; Humans; Learning algorithms; Liver; Liver cancer; Liver Neoplasms - diagnostic imaging; Liver Neoplasms - secondary; Machine learning; Male; Metastases; Metastasis; Middle Aged; Neural networks; Oncology; Optimization; Pilot Projects; Prediction models; Radiomics; Research Paper; Robustness; Segmentation; Surgical Oncology; Tomography, X-Ray Computed - methods; Variation; Deep learning; Machine learning; Biomarkers; Tomography; X-ray; Liver neoplasms; Computed
• DOI: 10.1007/s10585-021-10119-6
• ISSN: 0262-0898

Histopathological growth patterns (HGPs) are independent prognosticators for colorectal liver metastases (CRLM). Currently, HGPs are determined postoperatively. In this study, we evaluated radiomics for preoperative prediction of HGPs on computed tomography (CT), and its robustness to segmentation and acquisition variations. Patients with pure HGPs [i.e. 100% desmoplastic (dHGP) or 100% replacement (rHGP)] and a CT-scan who were surgically treated at the Erasmus MC between 2003–2015 were included retrospectively. Each lesion was segmented by three clinicians and a convolutional neural network (CNN). A prediction model was created using 564 radiomics features and a combination of machine learning approaches by training on the clinician’s and testing on the unseen CNN segmentations. The intra-class correlation coefficient (ICC) was used to select features robust to segmentation variations; ComBat was used to harmonize for acquisition variations. Evaluation was performed through a 100 × random-split cross-validation. The study included 93 CRLM in 76 patients (48% dHGP; 52% rHGP). Despite substantial differences between the segmentations of the three clinicians and the CNN, the radiomics model had a mean area under the curve of 0.69. ICC-based feature selection or ComBat yielded no improvement. Concluding, the combination of a CNN for segmentation and radiomics for classification has potential for automatically distinguishing dHGPs from rHGP, and is robust to segmentation and acquisition variations. Pending further optimization, including extension to mixed HGPs, our model may serve as a preoperative addition to postoperative HGP assessment, enabling further exploitation of HGPs as a biomarker.