Computer-Aided Detection, Segmentation and Characterization of Tumors in 3D medical images
The aim is our computer-aided tools is to improve the workflow of the radiologist when performing the tasks that are required by our proprietary method for the prediction of the resistance to NACT:
-1- Fast tumor localization: thanks to our localization tool, the radiologist can rapidly locate the tumor in MRI/PET images.
-2- Fast tumor 3D segmentation:
Medical image segmentation is a difficult clinical problem that will never be perfectly solved by computers, according to radiology experts. Furthermore, the accurate segmentation of lesions in MRI & PET images is generally time-consuming, and demands up to 15-30 minutes per case.
Hence, there is a real need for a fast-interactive delineation method, the aim being to improve the workflow of the radiologist. To this aim, we have designed a proprietary semi-automatic 3D segmentation algorithm that integrates state-of-the-art computer vision algorithms.
Our semi-automatic segmentation method has been successfully applied to head-neck lesions in 3D PET images. Most of the automatic segmentations were correct, only a few of them were interactively updated. As a result, the average total time taken by the user to complete the 3D tumor segmentation was about 1 minute per case, instead of 15-30 minutes.
-3- Tumor functional or metabolic characterization: radiomic predictive features generated by multidimensional analysis of multi-sequence/multimodality imaging data are used to generate the training data of our method for the early prediction of the resistance to NACT in Cancer.
Early Prediction of the Resistance to Neoadjuvant Chemotherapy in Breast Cancer
NACT for Breast Cancer: Breast cancer is the most common cancer in women worldwide and the second most common cancer overall (BRCF 2017-2018). Early breast cancer is traditionally treated with surgery, plus radiotherapy and adjuvant systemic therapy as required (Vaidya 2018). Neoadjuvant chemotherapy (NACT) for locally advanced breast cancer is a strategy that was introduced two decades ago to decrease the tumor size in large cancers and enable breast-conservative treatment (Vaidya 2018). A majority of patients are resistant to NACT treatment. Avoiding unnecessary treatments that can have heavy undesirable side effects and might be cardiotoxic is particularly important to patients since it would improve their quality of life. Furthermore, economic savings might be obtained by avoiding unnecessary and costly treatments. This value for money aspect is important for cost-sensitive healthcare funders.
Prediction of the Response to NACT: We have designed a reliable AI-powered method for the early prediction of response to NACT, by analyzing jointly predictive biomarkers and multisequence/multimodality imaging data (DCE-MRI, DWI, and PET). The prediction is performed at the baseline, i.e. before therapy beginning. A high specificity threshold has been set to get a reliable prediction by design.
Preliminary Results: The proof of concept has been performed using retrospective data from patients treated with NACT and classified into four cancer types: Luminal-A, Luminal-B, Triple Negative, and HER2+. About 75% of all non-responders have been reliably identified at baseline using retrospective data. The worst prediction score was obtained in the case of Triple-Negative breast cancers. The best prediction score is obtained for Luminal-A, Luminal-B, and HER2+ breast cancers, for which about 90% of non-responders are reliably detected at baseline.
Further possible improvements:
- Prediction of the response to NACT performed not only at baseline but also after the first cycle of NACT.
- Addition of multi-omics data and 3D whole breast ultrasound imaging data to improve the performance of the prediction method, particularly in the case of triple negative breast cancers.
Large-scale Multicentric clinical validation
The objective of the clinical study is to demonstrate the replicability of the results obtained by the proof of concept. If our results were confirmed by a large scale multi-centric clinical evaluation, unnecessary treatments could be stopped at an early stage. The expected economic impact would be particularly significant since about 55% of HER2+ patients do not respond to the treatment with Trastuzumab, which costs about €27,000 per patient for a full course of treatment.
References
[BRCF 2017-2018] Breast Cancer Research Foundation, Breast Cancer Statistics, https://www.bcrf.org/breast-cancer-statistics
[Vaidya 2018] Rethinking neoadjuvant chemotherapy for breast cancer, J.S. Vaidya et al., BMJ 2018, https://www.bmj.com/content/360/bmj.j5913.abstract/Response