2022-10-004 - DeepRDT - Development of Deep Learning Models to Predict Therapeutic Response of Radiotherapy in Cancer Patients

DeepRDT

2022-10-004

AI tool based on the analysis of medical images using deep learning (DL) models to predict the response to radiotherapy in patients with rectal and lung cancer. It should work as a guide for oncologists in the personalization of therapy

Social: health and safety

This proposal is at an early stage of the innovative product development path but is susceptible to have great impact, and opens the use of this novel technology (image analysis with deep neural networks) to other applications in the field. The expertise gained in this proposal will foster additional application of the analysis of images on other cancers or areas. The possibility to generate these products depends on the availability of data, images and clinical annotations, something that has been managed to obtain in a large sample size. This project has high potential for translation to the clinics and derives from needs requested by clinicians. The imaging analytical tools that are being developed have ambitious aims, that will assist oncologists and radiologists in their daily work to better care patients, with personalized strategies. The models applied to image analysis will help to identify where the tumor is, to quantify its volume and evolution with treatment, and to provide an early prediction of prognosis and potential toxicity – the latter being especially relevant in lung cancer, where radiation therapy often results in pneumonitis or fibrosis. Identifying patients with tumors that might be resistant to radiotherapy will allow the oncologist designing alternative strategies.

Artificial intelligence (AI) tool based on the analysis of medical images using deep learning (DL) models to predict response to radiation therapy in patients with rectal and lung cancer, serving as a guide for oncologists in personalizing therapy. Two scenarios have been selected for the proof of concept: a) early-stage lung cancer treated with stereotactic body radiation therapy (SBRT) and b) localized rectal cancer treated with neoadjuvant radiation therapy. The methods are applicable to other types of cancers treated with radiation therapy. Deep convolutional neural network models will be used to extract relevant information from diagnostic computed tomography. Data from positron emission tomography, magnetic resonance imaging, tumor biopsy histopathology, and clinical-molecular data will also be integrated when available.

0

0

0

The Business Development Plan on this project is to transfer it to a company for royalties of 3% -5% and set different milestones within the development plan (once it has been validated, after finishing the regulatory and when the project is "ready to sell", not deductible from royalties).

Radiation therapy has become a fundamental pillar for the treatment of multiple cancers, with proven benefits in treatment outcomes, both as neo/adjuvant therapy or as a standalone treatment. However, it is not exempt from adverse effects, and there is variability in efficacy among patients, with 20-40% of cases not responding to treatment, depending on tumor characteristics. T

Cancer patients treated with radiotherpy. Initially focusing on rectal and lung cancer.

Funding for further validation Codevelopment with a specialized company

Collaborations and Funding

Deep learning model training and validation Prototype development Regulatory strategy and implementation

Personalized treatment based on IA on the radiotherapy procedure of cancer patients.

Acces to relevant clincal data and harmonization of the data

1. Patient identification 2. Data acquisition and Data Management Plan (DMP). a. Clinical annotation. A database for the study has been designed (redcap) to organize the patient clinical data, including tumor characteristics, radiotherapy treatment modality and doses, and the endpoints: tumor response (RECIST) and toxicity (pneumonitis related to radiation, only for lung cancer patients). b. Radiology and nuclear medicine image data (CT and PET scans) will be downloaded from PACS. 3. Image data pre-processing: anonymization, standardization, segmentation, quality control. 4. Deep learning model training & internal validation: Models will use data as they are available. Initial models already have been started to train. This process requires a search for optimal model architecture and hyperparameters. 5. Explainability analysis: while models are developed, the features that are relevant for good model predictions will be analyzed in parallel. 6. Tool prototype development. To easy implementation in the clinic, a tool with web interface will be designed, so that the clinician can upload the images and relevant data and get a probability of treatment response for the patient. 7. Intellectual property protection. As soon as obtained validated predictive accuracies > 85%, which can be considered useful to implement in clinics, a patentability study of the model algorithms and predictive tool will be contracted. In any case, the tool will be registered. 8. Regulatory strategy. The software classification will be analyzed as well as the defined development pathway for the tool to reach the market according to recent EU regulations. Expert services will be contracted for this task.

Defining IPR Strategy