Purpose. To develop a deep reinforcement learning (DRL) agent to self-interact with the treatment planning system to automatically generate intensity modulated radiation therapy (IMRT) treatment plans for head-and-neck (HN) cancer with consistent organ-at-risk (OAR) sparing performance. Methods. With IRB approval, one hundred and twenty HN patients receiving IMRT were included. The DRL agent was trained with 20 patients. During each inverse optimization process, the intermediate dosimetric endpoints’ values, dose volume constraints’ values and structure objective function losses were collected as the DRL states. By adjusting the objective constraints as actions, the agent learned to seek optimal rewards by balancing OAR sparing and planning target volume (PTV) coverage. Reward computed from current dosimetric endpoints and clinical objectives were sent back to the agent to update action policy during model training. The trained agent was evaluated with the rest 100 patients. Results. The DRL agent was able to generate a clinically acceptable IMRT plan within 12.4 ± 3.1 min without human intervention. DRL plans showed lower PTV maximum dose (109.2%) compared to clinical plans (112.4%) (p < .05). Average median dose of left parotid, right parotid, oral cavity, larynx, pharynx of DRL plans were 15.6 Gy, 12.2 Gy, 25.7 Gy, 27.3 Gy and 32.1 Gy respectively, comparable to 17.1 Gy, 15.7 Gy, 24.4 Gy, 23.7 Gy and 35.5 Gy of corresponding clinical plans. The maximum dose of cord + 5 mm, brainstem and mandible were also comparable between the two groups. In addition, DRL plans demonstrated reduced variability, as evidenced by smaller 95% confidence intervals. The total MU of the DRL plans was 1611 vs 1870 (p < .05) of clinical plans. The results signaled the DRL’s consistent planning strategy compared to the planners’ occasional back-and-forth decision-making during planning. Conclusion. The proposed DRL agent is capable of efficiently generating HN IMRT plans with consistent quality.

The COVID-19 pandemic has posed a significant global public health threat with an escalating number of new cases and death toll daily. The early detection of COVID-related CXR abnormality potentially allows the early isolation of suspected cases. Chest X-Ray (CXR) is a fast and highly accessible imaging modality. Recently, a number of CXR-based AI models have been developed for the automated detection of COVID-19. However, most existing models are difficult to interpret due to the use of incomprehensible deep features in their models. Confronted with this, we developed an interpretable TSK fuzzy system in this study for COVID-19 detection using radiomics features extracted from CXR images. There are two main contributions. (1) When TSK fuzzy systems are applied to classification tasks, the commonly used binary label matrix of training samples is transformed into a soft one in order to learn a more discriminant transformation matrix and hence improve classification accuracy. (2) Based on the assumption that the samples in the same class should be kept as close as possible when they are transformed into the label space, the compactness class graph is introduced to avoid overfitting caused by label matrix relaxation. Our proposed model for a multi-categorical classification task (COVID-19 vs. No-Findings vs. Pneumonia) was evaluated using 600 CXR images from publicly available datasets and compared against five state-of-the-art AI models in aspects of classification accuracy. Experimental findings showed that our model achieved classification accuracy of over 83%, which is better than the state-of-the-art models, while maintaining high interpretability.

Objective. Artificial intelligence (AI) based treatment planning tools are being implemented in clinic. However, human interactions with such AI tools are rarely analyzed. This study aims to comprehend human planner’s interaction with the AI planning tool and incorporate the analysis to improve the existing AI tool. Approach. An in-house AI tool for whole breast radiation therapy planning was deployed in our institution since 2019, among which 522 patients were included in this study. The AI tool automatically generates fluence maps of the tangential beams to create an AI plan. Human planner makes fluence edits deemed necessary and after attending physician approval for treatment, it is recorded as final plan. Manual modification value maps were collected, which is the difference between the AI-plan and the final plan. Subsequently, a human-AI interaction (HAI) model using full scale connected U-Net was trained to learn such interactions and perform plan enhancements. The trained HAI model automatically modifies the AI plan to generate AI-modified plans (AI-m plan), simulating human editing. Its performance was evaluated against original AI-plan and final plan. Main results. AI-m plan showed statistically significant improvement in hotspot control over the AI plan, with an average of 25.2cc volume reduction in breast V105% (p = 0.011) and 0.805% decrease in Dmax (p < .001). It also maintained the same planning target volume (PTV) coverage as the final plan, demonstrating the model has captured the clinic focus of improving PTV hot spots without degrading coverage. Significance. The proposed HAI model has demonstrated capability of further enhancing the AI plan via modeling human-AI tool interactions. This study shows analysis of human interaction with the AI planning tool is a significant step to improve the AI tool.

Purpose. Traditional machine learning (ML) and deep learning (DL) applications in treatment planning rely on complex model architectures and large, high-quality training datasets. However, they cannot fully replace the conventional optimization process. This study presents a novel application of ML in treatment planning where established ML/DL toolkits are directly applied to treatment plan optimization. Materials and Methods. A one-layer network was designed based on the dose deposition matrix and implemented in PyTorch’s L-BFGS optimizer with GPU acceleration. The classical steepest descent optimizer was selected as a reference for comparison. Both optimizers utilized identical inputs and objective functions to ensure a fair comparison. DVH- and gEUD-based objectives were implemented in standard quadratic forms. Standard uniform and 1,000 random initializations were used to test optimizer’s search ability under different starting conditions for prostate and head-and-neck cases. Results. The MLT-assisted framework demonstrated comparable or superior plan quality to classical optimization by achieving lower objective values, improved DVHs and capturing finer modulation details in fluence maps. For gEUD-based optimization, it effectively explored beam weight elevations that classical optimization could only reach with stricter convergence criteria and many more iterations. The quality differences primarily stemmed from convergence speed. The MLT-assisted framework required significantly fewer evaluations and iterations to achieve similar or better results. Optimization on random initial maps further demonstrated that it was more robust and less likely to be trapped. It does not require stricter convergence criteria or extended runs to reach high-quality optima, making it more efficient and reliable. Conclusion. This framework leverages ML toolkits in a novel way, enabling faster convergence, greater robustness and handling of complex constraints. As the first study of its kind, it establishes MLT-assisted optimization as a viable and effective alternative to classical methods.

In this study, we present an innovative approach that harnesses deep neural networks to simulate respiratory lung motion and extract local functional information from single-phase chest X-rays, thus providing valuable auxiliary data for early diagnosis of lung cancer. A novel radiograph motion simulation (RMS) network was developed by combining a U-Net and a long short-term memory (LSTM) network for image generation and sequential prediction. By utilizing a spatial transformer network to deform input images, our proposed network ensures accurate image generation. We conducted both qualitative and quantitative assessments to evaluate the effectiveness and accuracy of our proposed network. The simulated respiratory motion closely aligns with pulmonary biomechanics and reveals enhanced details of pulmonary diseases. The proposed network demonstrates precise prediction of respiratory motion in the test cases, achieving remarkable average Dice scores exceeding 0.96 across all phases. The maximum variation in lung length prediction was observed during the end-exhale phase, with average deviation of 4.76 mm (±6.64) for the left lung and 4.77 mm (±7.00) for the right lung. This research validates the feasibility of generating patient-specific respiratory motion profiles from single-phase chest radiographs.