ADVERTISEMENT
Differentiating subcentimeter lung metastases in colorectal cancer patients by radiomics and deep learning approaches: A multicenter study
Preoperative evaluation of the indeterminate lung nodules is important for individual treatment of colorectal cancer (CRC). We aim to develop and validate discriminative radiomics and deep learning approaches for differentiating subcentimenter lung metastases (LMs) in CRC patients.
Models were developed in a primary cohort included 1194 consecutive CRC patients with initial subcentimeter size LMs on CT. Patients were randomly assigned (7:3) to the training or internal validation cohorts (IVC). Machine learning (ML), deep learning (DL), and the integration of radiomics and DL features were applied to classify the subcentimeter lung nodules as LMs or benign lesions. Two independent external validation cohorts (EVC) consisted of 101 (EVC1) and 40 (EVC2) patients. To verify the generalizability for nodules of smaller sizes, stepwise validations on the subgroups according to the nodule’s largest diameter (10, 9, 8, 7, 6, 5, ≤4 mm) were conducted.
The diagnostic accuracy by radiologists was 0.705 in primary cohort. The best ML based on support vector machine (SVM) showed a 0.981 area under the curve (AUC) in IVC, a 0.961 and 0.996 AUC in EVC1 and EVC2. The best integration model showed a 0.973 AUC in the IVC and a 0.943 and 0.974 AUC in EVC1 and EVC2. The DL model showed a 0.953 AUC in the IVC and a 0.907 and 0.951 AUC in EVC1 and EVC2. Stepwise validation demonstrated that with the LM diameter decreasing, the integration model was the most stable with smaller LMs, and was the best for LM ≤5 mm.
Our findings showed that radiomics and deep learning approaches based on CT could improve current diagnostic accuracy of subcentimeter CRC LMs. Specially, novel integration model was the best with LMs ≤5 mm. This study provided an automatic and noninvasive solution for determining subcentimeter LMs in the individual management of CRC.
The authors.
Has not received any funding.
All authors have declared no conflicts of interest.
