Deep learning‐based radiomics predicts response to chemotherapy in colorectal liver metastases

Abstract

Purpose The purpose of this study was to develop and validate a deep learning (DL)‐based radiomics model to predict the response to chemotherapy in colorectal liver metastases (CRLM). Methods In this retrospective study, we enrolled 192 patients diagnosed with CRLM who received first‐line chemotherapy and were followed by response assessment. Tumor response was identified according to the Response Evaluation Criteria in Solid Tumors (RECIST). Contrast‐enhanced multidetector computed tomography (MDCT) images were fed as inputs of the ResNet10‐based DL radiomics model, and the possibility of response was predicted as the output. The final combined DL radiomics model was constructed by integrating the response‐related clinical factors and the developed DL radiomics signature. A time‐independent validation cohort (n = 48) was extracted from the 192 patients to evaluate the DL model with area under the receiver operating characteristic curve (AUC), specificity, and sensitivity. Meanwhile, a traditional radiomics model was constructed using least absolute shrinkage and selection operator (lasso) as comparisons with the DL‐based model. Results According to RECIST criteria, 131 patients were identified as responders with complete response, partial response, and stable disease, while 61 patients were nonresponders with progression disease. The selected predictive clinical factor turned out to be the carcinoembryonic antigen (CEA) level with AUC of 0.489 (95% confidence interval [CI], 0.380–0.599) and 0.558 (95% CI, 0.374–0.741) in the training and validation cohorts, respectively. The DL‐based model provided better performance than the traditional classifier‐based radiomics model with significantly higher AUC (training: 0.903 [95% CI, 0.851–0.955] vs 0.745 [95% CI, 0.659–0.831]; validation: 0.820 [95% CI, 0.681–0.959] vs 0.598 [95% CI, 0.422–0.774]). The combination of DL‐based model with the CEA level provided slightly increased performance with AUC of 0.935 [95% CI, 0.897–0.973] in the training cohort and 0.830 [95% CI, 0.688‐0.973] in the validation cohort. Conclusions The developed DL‐based radiomics model could improve the efficiency to predict the response to chemotherapy in CRLM, which may assist in subsequent personalized treatment decision‐making in CRLM management.

Artificial intelligence in the diagnosis and management of colorectal cancer liver metastases

“…Discriminating responsive from unresponsive nodules or new lesions on the CT scan often represents a challenging task for radiologists, therefore Maaref et al [ 54 ] developed a fully automated framework based on DL CNN that achieved an accuracy of 0.91 (95%CI: 0.88-0.93) for differentiating treated and untreated lesions, and 0.78 (95%CI: 0.74-0.83) for predicting the response to a FOLFOX + bevacizumab-based chemotherapy regimen. Similarly, the DL radiomics model by Wei et al [ 55 ] was able to predict response to chemotherapy (CAPEOX, mFOLFOX6, FOLFIRI or XELIRI regimens) of CRLM based on contrast-enhanced CT according to the response evaluation criteria in solid tumors with an AUC in the validation cohort of 0.820 (95%CI: 0.681-0.959) that increases to 0.830 (95%CI: 0.688-0.973) combining the DL-based model with the CEA serum level. Human epidermal growth factor receptor 2 amplification or overexpression is found in 2%-6% of stage 2/3 CRC patients and treatment with trastuzumab and lapatinib has proven to be beneficial in the 70% of metastatic cases[ 56 ].…”

Is precision medicine for colorectal liver metastases still a utopia? New perspectives by modern biomarkers, radiomics, and artificial intelligence

“…This method was later validated against the multidisciplinary team meeting achieving an almost perfect agreement[ 108 ]. Since then, few studies have been published, but they outlined a progressive increase in AI performances[ 109 - 113 ]. The AI predicted the recurrence risk after surgery by taking into account clinical, pathology, and laboratory data[ 110 , 112 ].…”

“…The addition of radiomic features into the machine learning models further optimized and anticipated the prediction[ 109 , 111 ]. Wei et al [ 113 ] compared a clinical, radiomic, and AI-based model to predict response to first-line chemotherapy; the deep-learning model had the best results, outperforming not only the model based on clinical parameters but also the one including texture analysis.…”

Radiomics in hepatic metastasis by colorectal cancer

“…DL self-learning quantitative features may supplement unrevealed imaging features besides conventional radiomic features to improve the predictive power. Additionally, DL-based radiomics avoided time-consuming [ 23 ].…”