In this study, we evaluate the influence of normalization on the performance of deep learning networks for tumor segmentation and the prediction of the pathological response of locally advanced rectal cancer to neoadjuvant chemoradiotherapy. The techniques were applied to a multicenter and multimodal magnet resonance imaging data set consisting of 201 patients recorded at six centers. We implemented and investigated six different normalization methods (setting the mean and standard deviation, histogram matching, percentiles, combining percentiles and histogram matching, fixed window and an auto-encoder with adversarial loss using the imaging parameters) and evaluated their impact on four deep learning tasks: tumor segmentation, prediction of treatment outcome, and prediction of sex and age. The latter two tasks were implemented as a reference test. We trained a modified U-Net with different normalization methods in multiple configurations: on all images, images from all centers except one, and images from a single center. Our results show that normalization only plays a minor role in segmentation, with a difference in Dice of less than 0.02 between the best and worst performing networks. For the prediction of sex and treatment outcomes, the percentile method combined with histogram matching works best for all scenarios. The biggest difference in performance, depending on the normalization method, occurs for classification. In conclusion, normalization is especially important for small data sets or for generalizing to different data distributions. The deep learning method was superior to the classical methods only in a minority of cases, probably due to the limited amount of training data.

This retrospective study aims to evaluate the generalizability of a promising state-of-the-art multitask deep learning (DL) model for predicting the response of locally advanced rectal cancer (LARC) to neoadjuvant chemoradiotherapy (nCRT) using a multicenter dataset. To this end, we retrained and validated a Siamese network with two U-Nets joined at multiple layers using pre- and post-therapeutic T2-weighted (T2w), diffusion-weighted (DW) images and apparent diffusion coefficient (ADC) maps of 83 LARC patients acquired under study conditions at four different medical centers. To assess the predictive performance of the model, the trained network was then applied to an external clinical routine dataset of 46 LARC patients imaged without study conditions. The training and test datasets differed significantly in terms of their composition, e.g., T-/N-staging, the time interval between initial staging/nCRT/re-staging and surgery, as well as with respect to acquisition parameters, such as resolution, echo/repetition time, flip angle and field strength. We found that even after dedicated data pre-processing, the predictive performance dropped significantly in this multicenter setting compared to a previously published single- or two-center setting. Testing the network on the external clinical routine dataset yielded an area under the receiver operating characteristic curve of 0.54 (95% confidence interval [CI]: 0.41, 0.65), when using only pre- and post-therapeutic T2w images as input, and 0.60 (95% CI: 0.48, 0.71), when using the combination of pre- and post-therapeutic T2w, DW images, and ADC maps as input. Our study highlights the importance of data quality and harmonization in clinical trials using machine learning. Only in a joint, cross-center effort, involving a multidisciplinary team can we generate large enough curated and annotated datasets and develop the necessary pre-processing pipelines for data harmonization to successfully apply DL models clinically.

The purpose of the present development is to employ 3D printing to prototype an ion chamber array with a scalable design potentially allowing increased spatial resolution and a larger active area. An additional goal is to design and fabricate a custom size thin-panel detector array with low-Z components. As a proof of principle demonstration, a medium size detector array with 30 × 30 air-vented ion chambers was 3D-printed using PLA as frame for the electrodes. The active-area is 122 mm × 120 mm with 4 × 4 mm ² spatial resolution. External electrodes are cylindrical and made from conductive PLA. Internal electrodes are made from microwire. The array is symmetric with respect to the central plane and its thickness is 10 mm including build-up/-down plates of 2.5 mm thickness. Data acquisition is realized by biasing only selected chamber rows and reading only 30 chambers at a time. To test the device for potential clinical applications, 1D dose profiles and 2D dose maps with various square and irregular fields were measured. The overall agreement with the reference doses (film and treatment planning system) was satisfactory, but the measured dose differs in the penumbra region and in the field size dependence. Both of these features are related to the thin walls between neighboring ion chambers and different lateral phantom scatter in the detector panel vs homogeneous material. We demonstrated feasibility of radiation detector arrays with minimal number of readout channels and low-cost electronics. The acquisition scheme based on selected row or column ‘activation’ by bias voltage is not practical for 2D dosimetry but it allows for rapid turn-around when testing of custom arrays with the aid of multiple 1D dose profiles. Future progress in this area includes overcoming the limitations due high chamber packing ratio, which leads to the lateral scattering effects.

Group-equivariant convolutional neural networks (G-CNN) heavily rely on parameter sharing to increase CNN's data efficiency and performance. However, the parameter-sharing strategy greatly increases the computational burden for each added parameter, which hampers its application to deep neural network models. In this paper, we address these problems by proposing a non-parameter-sharing approach for group equivariant neural networks. The proposed methods adaptively aggregate a diverse range of filters by a weighted sum of stochastically augmented decomposed filters. We give theoretical proof about how the group equivariance can be achieved by our methods. Our method applies to both continuous and discrete groups, where the augmentation is implemented using Monte Carlo sampling and bootstrap resampling, respectively. Our methods also serve as an efficient extension of standard CNN. The experiments show that our method outperforms parameter-sharing group equivariant networks and enhances the performance of standard CNNs in image classification and denoising tasks, by using suitable filter bases to build efficient lightweight networks.