Medical imaging is considered one of the most important advances in the history of medicine and has become an essential part of the diagnosis and treatment of patients. Earlier prediction and treatment have been driving the acquisition of higher image resolutions as well as the fusion of different modalities, raising the need for sophisticated hardware and software systems for medical image registration, storage, analysis, and processing. In this scenario and given the new clinical pipelines and the huge clinical burden of hospitals, these systems are often required to provide both highly accurate and real-time processing of large amounts of imaging data. Additionally, lowering the prices of each part of imaging equipment, as well as its development and implementation, and increasing their lifespan is crucial to minimize the cost and lead to more accessible healthcare. This paper focuses on the evolution and the application of different hardware architectures (namely, CPU, GPU, DSP, FPGA, and ASIC) in medical imaging through various specific examples and discussing different options depending on the specific application. The main purpose is to provide a general introduction to hardware acceleration techniques for medical imaging researchers and developers who need to accelerate their implementations.

This paper provides an overview of the different deep convolutional neural network (DCNNs) architectures that have been investigated in the past years for the generation of synthetic computed tomography (CT) or pseudo-CT from magnetic resonance (MR). The U-net, the Atrous-net and the Residual-net architectures were analyzed, implemented and compared. Each network was implemented using 2D filters and 3D filters with 2D slices and 3D patches respectively as inputs. Two datasets were used for training and evaluation. The first one is composed by pairs of 3D T1-weighted MR and Low-dose CT images from the head of 19 healthy women. The second database contains dual echo Dixon-VIBE MR images and CT images from the pelvis of 13 colorectal and 6 prostate cancer patients. Bone structures in the target anatomy were key in choosing the right deep learning approach. This work provides a deep explanation of the architectures in order to know which DCNN fits better each medical application. According to this study, the 3D U-net architecture would be the best option to generate head pseudo-CTs while the 2D Residual-net provides the most accurate results for the pelvis anatomy.

Accurately estimating relapse risk in localized colon cancer (LCC) remains a challenge, as clinicopathological staging often fails to differentiate patients with a higher likelihood of recurrence. There is a need for novel tools to improve patient selection for post-operative chemotherapy. Radiomics has emerged as a powerful, noninvasive approach that may enhance clinical decision making. This retrospective study selected consecutive stage II and III LCC patients operated with curative intent from 2015 to 2017 in two academic institutions. Patients were assigned to either a training cohort made up of 80% of them or a test cohort, to further validate the initial findings. Penalized Cox proportional hazards and gradient boosted algorithms were designed to estimate time to relapse following a five-fold cross-validation process. Three models were assessed: (i) based only on clinical and pathological features, (ii) on radiomic features alone, and (iii) including clinical/pathological and radiomic variables. A new 'Risk Classification' score was generated based on the best risk assessment. A total of 278 patients were included in both cohorts. The Cox model trained with clinical and imaging variables showed the highest prognostic power, with a C-index of 0.68 and a mean cumulative dynamic area under the curve (AUC) of 0.69 on the test set. Feature screening identified 20 variables, including clinical data, radiomics features, and fractal features. SHapley Additive exPlanations (SHAP) analysis highlighted factors related to geometry, vascular invasion, and tumor stage as significant variables related to relapse. The new 'Risk Classification' score was able to identify patients with high risk of relapse both in univariable [hazard ratio (HR) 14.22, 95% confidence interval (CI) 1.91-106.08, P = 0.010] and multivariable (HR 11.74, 95% CI, 1.54-89.34, P = 0.017) models. Risk analysis revealed the new 'Risk Classification' variable as the one with the highest prognostic power compared with the ones currently used. Our findings suggest the potential for improved time-to-relapse estimation, enabling better patient stratification.

Objectives: The aim of this study was to develop and validate an MRI radiomics-based predictive model to discriminate significant prostate cancer (sPCa), compare it with PI-RADS, and determine whether incorporating PI-RADS and other clinical variables improves clinical performance. Materials and methods: A retrospective observational study was conducted using a cohort of 1497 MRI cases from 1395 men to develop the models. For each case, the index-lesion PI-RADS score, systematic ± targeted biopsy results, and six additional clinical variables were collected. Prostate biopsy samples served as the reference standard, defining sPCa as Gleason Grade ≥ 7. Handcrafted radiomic features were extracted from automatically segmented prostate glands. Four machine learning models were developed: (1) Radiomics, (2) PI-RADS, (3) PI-RADS + Radiomics, and (4) PI-RADS + Radiomics + Clinical Variables. Model performance and comparisons were evaluated using the area under the curve (AUC), while clinical utility was assessed through the decision curve analysis plot, Clinical Utility plot, and the number of avoided biopsies. Results: The radiomics model did not perform significantly better than PI-RADS in the validation cohort (AUC 0.838 vs. 0.833, p = 0.874). The combination of radiomics, PI-RADS, and clinical variables achieved the highest performance, with an AUC of 0.891 (95% CI: 0.853–0.930), significantly outperforming the other models ( p < 0.05). It also showed the highest specificity (29.41%) and biopsy avoidance rate (18.15%), although the differences were not statistically significant ( p = 0.313). Conclusions: Incorporating radiomics and clinical variables into PI-RADS enhances its ability to discriminate sPCa, potentially decreasing false positives and unnecessary biopsies. Critical relevance statement: The incorporation of clinical variables and radiomics into PI-RADS enhances its ability to predict significant prostate cancer, helping mitigate some of PI-RADS’s current limitations, such as a significant false-positive rate, and might help reduce unnecessary biopsies. Key Points: PI-RADS limitations result in overdiagnosis of indolent prostatic lesions and unnecessary biopsies. Radiomics and clinical variables enhance PI-RADS ability to detect significant prostate cancer. Combined clinical-radiological models reduce false positives and help avoid unnecessary biopsies.