HCC Detection in the Wild: AI models for automated MRI HCC Detection

Purpose: To develop and assess the effectiveness of artificial intelligence models to detect (localize) hepatocellular carcinoma (HCC) tumors from volumetric magnetic resonance (MR) images.

Materials and Methods: A total of 353 patients with a single HCC tumor who underwent transarterial radioembolization between 1/2004 and 3/2017 were identified. The mean age was 66 years with a standard deviation of 10.3 years. The cohort included 66% males. Pre-procedure MRI studies were acquired from the vendor neutral archive (VNA). A total of 249 cases were included in the analysis; 30 patients did not have MRI images, 11 cases were missing from the VNA, and 63 cases were missing sequences or had too much motion to identify the lesion. Tumors ranged in volume from 0.5 cc to 2997 cc with a mean of 89 cc. All livers and tumors were manually segmented by an interventional radiology resident. A synergynet {1} model was then trained and used to segment the livers in 3D. The AI segmented liver volumes were then cropped and resized to 128 x 128 x 64 voxel volumes and used to train detection models. Detection was performed in 3D. Augmentation including affine transforms, scaling, motion injection, spike injection, gamma adjustment among others were used during training. A total of 198 cases were used for training, 25 for validation and 26 for testing. Detection quality was measured by average precision (AP) at intersection over union thresholds (IoU) of 0.1, 0.5 and 0.8.

Results: Three models were trained and tested on the same cases. The strongest performing model was the RetinaNet {2} detector with a ResNet {3} backbone which had AP performance of 94%, 72% and 41% at IoU thresholds of 0.1, 0.5 and 0.8. The nnDetect {4} model, which is an excellent baseline detection model for many problems had performance of 91%, 82% and 4%. We also trained a RetinaNet detector with a Darknet19 {5}  backbone (same backbone used in the YOLO family of detection models) had performance of 94%, 48% and 0% across the IoU thresholds. Model performance is summarized in the table.

Conclusion: This study demonstrates the feasibility of AI based 3D lesion detection models for localization of HCC lesions in MRI studies. The detection models performed well across a wide array of training samples with varying imaging characteristics and tumor sizes. Such AI models may play a role in future treatment planning software as well as diagnostic tasks.