In oncology, texture features extracted from positron emission tomography with 18-fluorodeoxyglucose images (FDG-PET) are of increasing interest for predictive and prognostic studies, leading to several tens of features per tumor. To select the best features, the use of a random forest (RF) classifier was investigated. Sixty-five patients with an esophageal cancer treated with a combined chemo-radiation therapy were retrospectively included. All patients underwent a pretreatment whole-body FDG-PET. The patients were followed for 3 years after the end of the treatment. The response assessment was performed 1 month after the end of the therapy. Patients were classified as complete responders and non-complete responders. Sixty-one features were extracted from medical records and PET images. First, Spearman’s analysis was performed to eliminate correlated features. Then, the best predictive and prognostic subsets of features were selected using a RF algorithm. These results were compared to those obtained by a Mann-Whitney U test (predictive study) and a univariate Kaplan-Meier analysis (prognostic study). Among the 61 initial features, 28 were not correlated. From these 28 features, the best subset of complementary features found using the RF classifier to predict response was composed of 2 features: metabolic tumor volume (MTV) and homogeneity from the co-occurrence matrix. The corresponding predictive value (AUC = 0.836 ± 0.105, Se = 82 ± 9%, Sp = 91 ± 12%) was higher than the best predictive results found using the Mann-Whitney test: busyness from the gray level difference matrix (P < 0.0001, AUC = 0.810, Se = 66%, Sp = 88%). The best prognostic subset found using RF was composed of 3 features: MTV and 2 clinical features (WHO status and nutritional risk index) (AUC = 0.822 ± 0.059, Se = 79 ± 9%, Sp = 95 ± 6%), while no feature was significantly prognostic according to the Kaplan-Meier analysis. The RF classifier can improve predictive and prognostic values compared to the Mann-Whitney U test and the univariate Kaplan-Meier survival analysis when applied to several tens of features in a limited patient database.

Distributed learning across coalitions is becoming popular for multi-centric implementation of deep learning models. However, the level of trust between the members of a coalition can vary and requires different security architectures. Privacy of the training data has been largely described in distributed learning. In this paper, we present a scalable security architecture providing additional features such as validation on the sources quality, confidentiality on the model within a trusted coalition or confidentiality among untrusted partners inside the coalition. More specifically, we propose solutions that guarantee preservation not only of data privacy but also of data quality, enforce a trustworthy sequence of iterative learning, and that lead to equitable sharing of the learned model among the coalition’s members. We give an example of its deployment in the case of the distributed optimization of a deep learning convolutional neural network trained on medical images.

External beam radiotherapy cancer treatment aims to deliver dose fractions to slowly destroy a tumor while avoiding severe side effects in surrounding healthy tissues. To automate the dose fraction schedules, this paper investigates how deep reinforcement learning approaches (based on deep Q network and deep deterministic policy gradient) can learn from a model of a mixture of tumor and healthy cells. A 2D tumor growth simulation is used to simulate radiation effects on tissues and thus training an agent to automatically optimize dose fractionation. Results show that initiating treatment with large dose per fraction, and then gradually reducing it, is preferred to the standard approach of using a constant dose per fraction.

For prostate cancer patients, large organ deformations occurring between radiotherapy treatment sessions create uncertainty about the doses delivered to the tumor and surrounding healthy organs. Segmenting those regions on cone beam CT (CBCT) scans acquired on treatment day would reduce such uncertainties. In this work, a 3D U-net deep-learning architecture was trained to segment bladder, rectum, and prostate on CBCT scans. Due to the scarcity of contoured CBCT scans, the training set was augmented with CT scans already contoured in the current clinical workflow. Our network was then tested on 63 CBCT scans. The Dice similarity coefficient (DSC) increased significantly with the number of CBCT and CT scans in the training set, reaching $0.874$ ± $0.096$ , $0.814$ ± $0.055$ , and $0.758$ ± $0.101$ for bladder, rectum, and prostate, respectively. This was about 10% better than conventional approaches based on deformable image registration between planning CT and treatment CBCT scans, except for prostate. Interestingly, adding 74 CT scans to the CBCT training set allowed maintaining high DSCs, while halving the number of CBCT scans. Hence, our work showed that although CBCT scans included artifacts, cross-domain augmentation of the training set was effective and could rely on large datasets available for planning CT scans.

Objective and design: The objective of the present study was to assess the discriminative accuracy of artificial intelligence (AI) software to identify COPD and other chronic respiratory diseases from primary care spirometry. This was a diagnostic study with blinded analysis. Methods: Retrospective hand-held spirometry data from consecutive patients attending primary care clinics in Hillingdon (London, UK) between September 2015 and March 2019 were used. The index diagnosis was the “preferred” diagnosis determined by AI software (highest probability) using supervised random-forest machine learning to interpret raw spirometry data and basic demographics. The reference diagnosis was based on the consensus of expert pulmonologists with access to primary and secondary care medical notes and results of relevant investigations. Cross-tabulation of the index test results by the results of the reference standard for COPD and other respiratory disease categories provided the main outcome measures. Results: In this primary care spirometry dataset from 1113 patients, 543 (48.8%) had a reference diagnosis of COPD. AI preferred diagnosis detected 456, achieving a sensitivity of 84.0% (95% CI 80.6–87.0%), specificity of 86.8% (83.8–89.5%), accuracy of 85.4% (83.2–87.5%) with area under curve (AUC) of 0.914 (0.896–0.930). AI preferred diagnosis identified 187 out of 249 patients with reference diagnosis of interstitial lung disease and 59 out of 107 patients with asthma, with AUCs of 0.900 (0.880–0.916) and 0.814 (0.790–0.836), respectively. Conclusion: AI software achieved high sensitivity and specificity in identifying COPD using spirometry and basic demographic data and may support accurate diagnosis of COPD in primary care. AI software performed less well for other chronic respiratory disease categories.