Background Hypofractionated stereotactic radiotherapy (hFSRT) is a salvage option for recurrent glioblastoma (GB) which may synergize anti-PDL1 treatment. This phase I study evaluated the safety and the recommended phase II dose of anti-PDL1 durvalumab combined with hFSRT in patients with recurrent GB. Methods Patients were treated with 24 Gy, 8 Gy per fraction on days 1, 3, and 5 combined with the first 1500 mg Durvalumab dose on day 5, followed by infusions q4weeks until progression or for a maximum of 12 months. A standard 3 + 3 Durvalumab dose de-escalation design was used. Longitudinal lymphocytes count, cytokines analyses on plasma samples, and magnetic resonance imaging (MRI) were collected. Results Six patients were included. One dose limiting toxicity, an immune-related grade 3 vestibular neuritis related to Durvalumab, was reported. Median progression-free interval (PFI) and overall survival (OS) were 2.3 and 16.7 months, respectively. Multi-modal deep learning-based analysis including MRI, cytokines, and lymphocytes/neutrophil ratio isolated the patients presenting pseudoprogression, the longest PFI and those with the longest OS, but statistical significance cannot be established considering phase I data only. Conclusion Combination of hFSRT and Durvalumab in recurrent GB was well tolerated in this phase I study. These encouraging results led to an ongoing randomized phase II. (ClinicalTrials.gov Identifier: NCT02866747).

Deep reinforcement learning policies, despite their outstanding efficiency in simulated visual control tasks, have shown disappointing ability to generalize across disturbances in the input training images. Changes in image statistics or distracting background elements are pitfalls that prevent generalization and real-world applicability of such control policies. We elaborate on the intuition that a good visual policy should be able to identify which pixels are important for its decision, and preserve this identification of important sources of information across images. This implies that training of a policy with small generalization gap should focus on such important pixels and ignore the others. This leads to the introduction of saliency-guided Q-networks (SGQN), a generic method for visual reinforcement learning, that is compatible with any value function learning method. SGQN vastly improves the generalization capability of Soft Actor-Critic agents and outperforms existing stateof-the-art methods on the Deepmind Control Generalization benchmark, setting a new reference in terms of training efficiency, generalization gap, and policy interpretability.