Deep learning‐assisted magnetic resonance imaging prediction of tumor response to chemotherapy in patients with colorectal liver metastases

Abstract

Accurate evaluation of tumor response to preoperative chemotherapy is crucial for assigning appropriate patients with colorectal liver metastases (CRLM) to surgery or conservative therapy. However, there is no well‐recognized method for predicting pathological response before surgery. Our study constructed and validated a deep learning algorithm using prechemotherapy and postchemotherapy magnetic resonance imaging (MRI) to predict pathological response in CRLM. CRLM patients from center one who had ≤5 lesions and were scheduled to receive preoperative chemotherapy followed by liver resection between January 2013 and November 2016, were included prospectively and chronologically divided into a training cohort (80% of patients) and a testing cohort (20% of patients). Patients from center two were included January 2017 and December 2018 as an external validation cohort. MRI‐based models were constructed to discriminate according to pathology tumor regression grade (TRG) between the response (TRG1/2) and nonresponse (TRG3/4/5) groups at the lesion level. From center one, 155 patients (328 lesions) were included; chronologically, 101 (264 lesions) in the training cohort and 54 (64 lesions) in the testing cohort. The model achieved better accuracy (0.875 vs 0.578) and AUC (0.849 vs 0.615) than RECIST for discriminating response; it also distinguished the survival outcomes after hepatectomy better than the RECIST criteria. Evaluations of the external validation cohort (25 patients, 61 lesions) also showed good ability with an AUC of 0.833. In conclusion, the MRI‐based deep learning model provided accurate prediction of pathological tumor response to preoperative chemotherapy in patients with CRLM and may inform individualized treatment.

Deep learning for the prediction of early on-treatment response in metastatic colorectal cancer from serial medical imaging

“…However, until now, there is seldom report on applying DL method on predicting early on-treatment response in oncology, especially mCRC. For example, when searching in the National Library of Medicine medical literature database with the key words of 'deep learning', 'metastatic colorectal cancer' and 'response or survival' (https:// pubmed.ncbi.nlm.nih.gov/), only one literature was found 19 and the literature was only about colorectal liver metastases with small data. The challenges of predicting mCRC treatment response by using DL method include, (1) mCRC involves multiple metastatic lesions involving multiple organs, most commonly, liver, lung and lymph nodes; (2) the prediction of response involves CT images of multiple time points while the patient is on treatment rather than a single time point; and (3) DL method requires large dataset for training.…”

Deep Learning With 3D Convolutional Neural Network for Noninvasive Prediction of Microvascular Invasion in Hepatocellular Carcinoma

“…29 Deep learning has been widely applied in the field of liver cancer for diagnosis, treatment response, and prognosis. [30][31][32] Compared with radiomic analysis method, the deep learning method has a great advantage in learning from radiological images directly, instead of using human-defined features limited by human experiences. In this present study, we did not extract deep learning features but proposed a 3D CNN architecture that can identify MVI status end-to-end.…”

“…Deep learning has been widely applied in the field of liver cancer for diagnosis, treatment response, and prognosis 30–32 . Compared with radiomic analysis method, the deep learning method has a great advantage in learning from radiological images directly, instead of using human‐defined features limited by human experiences.…”

A Survey on Deep Learning for Precision Oncology

“…Zhu et al [ 82 ] proposed a densely connected center cropping CNN (DC3CNN) to predict chemotherapy response in patients with colorectal liver metastases by using pre-and post-chemotherapy MRI images. As shown in Figure 4 f, their architecture consists of four inputs, including pre-treatment T2-weighted image, pre-treatment apparent diffusion coefficient (ADC) map, post-treatment T2-weighted image, and post-treatment ADC map.…”