1209MO - Computer-aided HCC lesion detection based on deep learning and CT images

Background

Hepatocellular carcinoma (HCC) is the most prevalent type of primary liver cancer and a major cause of death worldwide. It can be identified using contrast-enhanced imaging, such as CT scans however the sensitivity of radiologists in detecting HCC lesions using CT scans remains low (Se=66% [60%-72%]). The proposed computer-aided diagnostic device has been developed based on deep learning and multiphase-CT scans, with the objective of assisting radiologists in improving the accuracy of HCC lesion detection.

Methods

The deep learning (DL) model is a combination of a specifically designed convolutional 3D classifier and a state-of-the-art 3D detector (3D-RetinaUnet), taking both arterial and portal phases as input. Upstream, the portal phase is registered on the arterial phase in a non-rigid fashion. The model outputs the location of HCC tumors as 3D bounding boxes. The dataset was collected at the Assistance Publique–Hôpitaux de Paris (APHP) from multiple sites (La Pitié-Salpêtrière, Beaujon and Paul Brousse) and annotated by at least 3 liver expert radiologists with the help of the sites' case report form as a guide.

Results

The dataset contained 820 multiphase CT with 966 tumors annotated by the radiologists. These 820 patients had not received treatment for HCC prior to the CT scan. The proposed DL model was trained on 629 multiphase-CT scans and tested on the multi-phase CT of 191 patients with 223 HCC lesions confirmed by histopathology. The sensitivity of HCC lesion diagnostic reached 89% (i.e., Lesion-level sensitivity) with a median of 2.0 false positives per scan (FPs/Scan) and a mean of 3.3 FPs/Scan. 69% of the 58 small-size HCC lesions measuring between 10 mm and 30 mm were diagnosed. The patient-level sensitivity was 96.3% which means the model identified the presence of HCC in 96.3% of the patients.

Conclusions

Our HCC diagnostic rate is above the means sensitivity of radiologists and our DL model could already participate to improve the diagnostic rate of HCC. The next goal is to improve the DL model and enrich the training set with patients at risk of HCC who are also treatment-naïve, including those with small HCC. Our objective is to develop an end-to-end CADe/CADx that will enhance the performance of detecting HCC, particularly small HCC.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Median Technologies.

Funding

Median Technologies.

Disclosure

S. Bodard, C. Bougharka: Financial Interests, Personal, Other, Consultancy: Median Technologies. L. Ayache, A. O'Keane, J. Shen, V.K. Le, E. Geremia, Q. Renault, M. Macrez, T. Foriel: Financial Interests, Personal, Full or part-time Employment: Median Technologies. S. Poullot: Financial Interests, Personal, Full or part-time Employment, Senior Data Scientist: Median Technologies; Financial Interests, Personal, Stocks/Shares: Median Technologies. V. Groza: Financial Interests, Personal, Advisory Board, Consulting: Median Technologies. Y. Liu: Financial Interests, Personal, Full or part-time Employment: Median Technologies; Financial Interests, Personal, Stocks/Shares: Median Technologies. V. Vilgrain: Financial Interests, Personal, Advisory Board: Guerbet, Guerbet, Sirtex, Siemens; Financial Interests, Personal, Invited Speaker: Sirtex, Canon Medical, GE Healthcare; Financial Interests, Institutional, Advisory Board: Quantum Chirurgical; Financial Interests, Institutional, Local PI, Industry/academic project funded by French government (Multi-industry company included GE Healthcare): HECAM, AIdream; Financial Interests, Institutional, Local PI, Industry/academic project funded by French government (Multi-industry company included Philips Medical System): RIHDO; Non-Financial Interests, Leadership Role, Scientific director of ESOR, no financial compensation: European School of Radiology (ESR). All other authors have declared no conflicts of interest.