Deep learning (DL) often requires an image quality metric; however, widely used metrics are not designed for medical images. To develop an image quality metric that is specific to MRI using radiologists image rankings and DL models. Retrospective. A total of 19,344 rankings on 2916 unique image pairs from the NYU fastMRI Initiative neuro database was used for the neural network-based image quality metrics training with an 80%/20% training/validation split and fivefold cross-validation. 1.5 T and 3 T T1, T1 postcontrast, T2, and FLuid Attenuated Inversion Recovery (FLAIR). Synthetically corrupted image pairs were ranked by radiologists (N = 7), with a subset also scoring images using a Likert scale (N = 2). DL models were trained to match rankings using two architectures (EfficientNet and IQ-Net) with and without reference image subtraction and compared to ranking based on mean squared error (MSE) and structural similarity (SSIM). Image quality assessing DL models were evaluated as alternatives to MSE and SSIM as optimization targets for DL denoising and reconstruction. Radiologists' agreement was assessed by a percentage metric and quadratic weighted Cohen's kappa. Ranking accuracies were compared using repeated measurements analysis of variance. Reconstruction models trained with IQ-Net score, MSE and SSIM were compared by paired t test. P < 0.05 was considered significant. Compared to direct Likert scoring, ranking produced a higher level of agreement between radiologists (70.4% vs. 25%). Image ranking was subjective with a high level of intraobserver agreement ( <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>94.9</mn> <mo>%</mo> <mo>±</mo> <mn>2.4</mn> <mo>%</mo></mrow> </math> ) and lower interobserver agreement ( <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>61.47</mn> <mo>%</mo> <mo>±</mo> <mn>5.51</mn> <mo>%</mo></mrow> </math> ). IQ-Net and EfficientNet accurately predicted rankings with a reference image ( <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>75.2</mn> <mo>%</mo> <mo>±</mo> <mn>1.3</mn> <mo>%</mo></mrow> </math> and <math xmlns="http://www.w3.org/1998/Math/MathML"><mrow><mn>79.2</mn> <mo>%</mo> <mo>±</mo> <mn>1.7</mn> <mo>%</mo></mrow> </math> ). However, EfficientNet resulted in images with artifacts and high MSE when used in denoising tasks while IQ-Net optimized networks performed well for both denoising and reconstruction tasks. Image quality networks can be trained from image ranking and used to optimize DL tasks. 3 TECHNICAL EFFICACY: Stage 1.

Various hepatocellular carcinoma (HCC) prediction models have been proposed for patients with chronic hepatitis B (CHB) using clinical variables. We aimed to develop an artificial intelligence (AI)-based HCC prediction model by incorporating imaging biomarkers derived from abdominal computed tomography (CT) images along with clinical variables. An AI prediction model employing a gradient-boosting machine algorithm was developed utilizing imaging biomarkers extracted by DeepFore, a deep learning-based CT auto-segmentation software. The derivation cohort (n = 5,585) was randomly divided into the training and internal validation sets at a 3:1 ratio. The external validation cohort included 2,883 patients. Six imaging biomarkers (i.e. abdominal visceral fat-total fat volume ratio, total fat-trunk volume ratio, spleen volume, liver volume, liver-spleen Hounsfield unit ratio, and muscle Hounsfield unit) and eight clinical variables were selected as the main variables of our model, PLAN-B-DF. In the internal validation set (median follow-up duration = 7.4 years), PLAN-B-DF demonstrated an excellent predictive performance with a c-index of 0.91 and good calibration function (p = 0.78 by the Hosmer-Lemeshow test). In the external validation cohort (median follow-up duration = 4.6 years), PLAN-B-DF showed a significantly better discrimination function compared to previous models, including PLAN-B, PAGE-B, modified PAGE-B, and CU-HCC (c-index, 0.89 vs. 0.65-0.78; all p <0.001), and maintained a good calibration function (p = 0.42 by the Hosmer-Lemeshow test). When patients were classified into four groups according to the risk probability calculated by PLAN-B-DF, the 10-year cumulative HCC incidence was 0.0%, 0.4%, 16.0%, and 46.2% in the minimal-, low-, intermediate-, and high-risk groups, respectively. This AI prediction model, integrating deep learning-based auto-segmentation of CT images, offers improved performance in predicting HCC risk among patients with CHB compared to previous models. The novel predictive model PLAN-B-DF, employing an automated computed tomography segmentation algorithm, significantly improves predictive accuracy and risk stratification for hepatocellular carcinoma in patients with chronic hepatitis B (CHB). Using a gradient-boosting algorithm and computed tomography metrics, such as visceral fat volume and myosteatosis, PLAN-B-DF outperforms previous models based solely on clinical and demographic data. This model not only shows a higher c-index compared to previous models, but also effectively classifies patients with CHB into different risk groups. This model uses machine learning to analyze the complex relationships among various risk factors contributing to hepatocellular carcinoma occurrence, thereby enabling more personalized surveillance for patients with CHB.

This study aims to evaluate a deep learning pipeline for detecting clinically significant prostate cancer (csPCa), defined as Gleason Grade Group (GGG) ≥ 2, using biparametric MRI (bpMRI) and compare its performance with radiological reading. The training dataset included 4381 bpMRI cases (3800 positive and 581 negative) across three continents, with 80% annotated using PI-RADS and 20% with Gleason Scores. The testing set comprised 328 cases from the PROSTATEx dataset, including 34% positive (GGG ≥ 2) and 66% negative cases. A 3D nnU-Net was trained on bpMRI for lesion detection, evaluated using histopathology-based annotations, and assessed with patient- and lesion-level metrics, along with lesion volume, and GGG. The algorithm was compared to non-expert radiologists using multi-parametric MRI (mpMRI). The model achieved an AUC of 0.83 (95% CI: 0.80, 0.87). Lesion-level sensitivity was 0.85 (95% CI: 0.82, 0.94) at 0.5 False Positives per volume (FP/volume) and 0.88 (95% CI: 0.79, 0.92) at 1 FP/volume. Average Precision was 0.55 (95% CI: 0.46, 0.64). The model showed over 0.90 sensitivity for lesions larger than 650 mm³ and exceeded 0.85 across GGGs. It had higher true positive rates (TPRs) than radiologists equivalent FP rates, achieving TPRs of 0.93 and 0.79 compared to radiologists' 0.87 and 0.68 for PI-RADS ≥ 3 and PI-RADS ≥ 4 lesions (p ≤ 0.05). The DL model showed strong performance in detecting csPCa on an independent test cohort, surpassing radiological interpretation and demonstrating AI's potential to improve diagnostic accuracy for non-expert radiologists. However, detecting small lesions remains challenging. Question Current prostate cancer detection methods often do not involve non-expert radiologists, highlighting the need for more accurate deep learning approaches using biparametric MRI. Findings Our model outperforms radiologists significantly, showing consistent performance across Gleason Grade Groups and for medium to large lesions. Clinical relevance This AI model improves prostate detection accuracy in prostate imaging, serves as a benchmark with reference performance on a public dataset, and offers public PI-RADS annotations, enhancing transparency and facilitating further research and development.

Deep learning-based segmentation of brain metastases relies on large amounts of fully annotated data by domain experts. Semi-supervised learning offers potential efficient methods to improve model performance without excessive annotation burden. This work tests the viability of semi-supervision for brain metastases segmentation. Retrospective. There were 156, 65, 324, and 200 labeled scans from four institutions and 519 unlabeled scans from a single institution. All subjects included in the study had diagnosed with brain metastases. 1.5 T and 3 T, 2D and 3D T1-weighted pre- and post-contrast, and fluid-attenuated inversion recovery (FLAIR). Three semi-supervision methods (mean teacher, cross-pseudo supervision, and interpolation consistency training) were adapted with the U-Net architecture. The three semi-supervised methods were compared to their respective supervised baseline on the full and half-sized training. Evaluation was performed on a multinational test set from four different institutions using 5-fold cross-validation. Method performance was evaluated by the following: the number of false-positive predictions, the number of true positive predictions, the 95th Hausdorff distance, and the Dice similarity coefficient (DSC). Significance was tested using a paired samples t test for a single fold, and across all folds within a given cohort. Semi-supervision outperformed the supervised baseline for all sites with the best-performing semi-supervised method achieved an on average DSC improvement of 6.3% ± 1.6%, 8.2% ± 3.8%, 8.6% ± 2.6%, and 15.4% ± 1.4%, when trained on half the dataset and 3.6% ± 0.7%, 2.0% ± 1.5%, 1.8% ± 5.7%, and 4.7% ± 1.7%, compared to the supervised baseline on four test cohorts. In addition, in three of four datasets, the semi-supervised training produced equal or better results than the supervised models trained on twice the labeled data. Semi-supervised learning allows for improved segmentation performance over the supervised baseline, and the improvement was particularly notable for independent external test sets when trained on small amounts of labeled data. Artificial intelligence requires extensive datasets with large amounts of annotated data from medical experts which can be difficult to acquire due to the large workload. To compensate for this, it is possible to utilize large amounts of un-annotated clinical data in addition to annotated data. However, this method has not been widely tested for the most common intracranial brain tumor, brain metastases. This study shows that this approach allows for data efficient deep learning models across multiple institutions with different clinical protocols and scanners. 3 TECHNICAL EFFICACY: Stage 2.

Immunotherapy has become an option for the first-line therapy of advanced gastric cancer (GC), with improved survival. Our study aimed to investigate unresectable GC from an imaging perspective combined with clinicopathological variables to identify patients who were most likely to benefit from immunotherapy. Patients with unresectable GC who were consecutively treated with immunotherapy at two different medical centers of Chinese PLA General Hospital were included and divided into the training and validation cohorts, respectively. A deep learning neural network, using a multimodal ensemble approach based on CT imaging data before immunotherapy, was trained in the training cohort to predict survival, and an internal validation cohort was constructed to select the optimal ensemble model. Data from another cohort were used for external validation. The area under the receiver operating characteristic curve was analyzed to evaluate performance in predicting survival. Detailed clinicopathological data and peripheral blood prior to immunotherapy were collected for each patient. Univariate and multivariable logistic regression analysis of imaging models and clinicopathological variables was also applied to identify the independent predictors of survival. A nomogram based on multivariable logistic regression was constructed. A total of 79 GC patients in the training cohort and 97 patients in the external validation cohort were enrolled in this study. A multi-model ensemble approach was applied to train a model to predict the 1-year survival of GC patients. Compared to individual models, the ensemble model showed improvement in performance metrics in both the internal and external validation cohorts. There was a significant difference in overall survival (OS) among patients with different imaging models based on the optimum cutoff score of 0.5 (HR = 0.20, 95 % CI: 0.10-0.37, <i>P</i> < 0.001). Multivariate Cox regression analysis revealed that the imaging models, PD-L1 expression, and lung immune prognostic index were independent prognostic factors for OS. We combined these variables and built a nomogram. The calibration curves showed that the C-index of the nomogram was 0.85 and 0.78 in the training and validation cohorts. The deep learning model in combination with several clinical factors showed predictive value for survival in patients with unresectable GC receiving immunotherapy.

Adjuvant radiotherapy is essential for reducing local recurrence and improving survival in breast cancer patients, but it carries a risk of ischemic cardiac toxicity, which increases with heart exposure. The isocentric lateral decubitus position, where the breast rests flat on a support, reduces heart exposure and leads to delivery of a more uniform dose. This position is particularly beneficial for patients with unique anatomies, such as those with pectus excavatum or larger breast sizes. While artificial intelligence (AI) algorithms for autocontouring have shown promise, they have not been tailored to this specific position. This study aimed to develop and evaluate a neural network-based autocontouring algorithm for patients treated in the isocentric lateral decubitus position. In this single-center study, 1189 breast cancer patients treated after breast-conserving surgery were included. Their simulation CT scans (1209 scans) were used to train and validate a neural network-based autocontouring algorithm (nnU-Net). Of these, 1087 scans were used for training, and 122 scans were reserved for validation. The algorithm's performance was assessed using the Dice similarity coefficient (DSC) to compare the automatically delineated volumes with manual contours. A clinical evaluation of the algorithm was performed on 30 additional patients, with contours rated by two expert radiation oncologists. The neural network-based algorithm achieved a segmentation time of approximately 4 min, compared to 20 min for manual segmentation. The DSC values for the validation cohort were 0.88 for the treated breast, 0.90 for the heart, 0.98 for the right lung, and 0.97 for the left lung. In the clinical evaluation, 90% of the automatically contoured breast volumes were rated as acceptable without corrections, while the remaining 10% required minor adjustments. All lung contours were accepted without corrections, and heart contours were rated as acceptable in 93.3% of cases, with minor corrections needed in 6.6% of cases. This neural network-based autocontouring algorithm offers a practical, time-saving solution for breast cancer radiotherapy planning in the isocentric lateral decubitus position. Its strong geometric performance, clinical acceptability, and significant time efficiency make it a valuable tool for modern radiotherapy practices, particularly in high-volume centers.

To address the challenges of verifying MR-based attenuation correction (MRAC) in PET/MR due to CT positional mismatches and alignment issues, this study utilized a flatbed insert and arms-down positioning during PET/CT scans to achieve precise MR-CT matching for accurate MRAC evaluation. A validation dataset of 21 patients underwent whole-body [¹⁸F]FDG PET/CT followed by [¹⁸F]FDG PET/MR. A flatbed insert ensured consistent positioning, allowing direct comparison of four MRAC methods-four-tissue and five-tissue models with discrete and continuous μ-maps-against CT-based attenuation correction (CTAC). A deep learning-based framework, trained on a dataset of 300 patients, was used to generate synthesized-CTs from MR images, forming the basis for all MRAC methods. Quantitative analyses were conducted at the whole-body, region of interest, and lesion levels, with lesion-distance analysis evaluating the impact of bone proximity on standardized uptake value (SUV) quantification. Distinct differences were observed among MRAC methods in spine and femur regions. Joint histogram analysis showed MRAC-4 (continuous μ-map) closely aligned with CTAC. Lesion-distance analysis revealed MRAC-4 minimized bone-induced SUV interference (r = 0.01, p = 0.8643). However, tissues prone to bone segmentation interference, such as the spine and liver, exhibited greater SUV variability and lower reproducibility in MRAC-4 compared to MRAC-2 (2D bone segmentation, discrete μ-map) and MRAC-3 (3D bone segmentation, discrete μ-map). Using a flatbed insert, this study validated MRAC with high precision. Continuous μ-value MRAC method (MRAC-4) demonstrated superior accuracy and minimized bone-related SUV errors but faced challenges in reproducibility, particularly in bone-rich regions.

Bi-parametric magnetic resonance imaging (bpMRI) has become a pivotal modality in the detection and diagnosis of clinically significant prostate cancer (csPCa). Developing AI-based systems to identify csPCa using bpMRI can transform prostate cancer (PCa) management by improving efficiency and cost-effectiveness. However, current state-of-the-art methods using convolutional neural networks (CNNs) and Transformers are limited in learning in-plane and three-dimensional spatial information from anisotropic bpMRI. Their performances also depend on the availability of large, diverse, and well-annotated bpMRI datasets. To address these challenges, we propose the Zonal-aware Self-supervised Mesh Network (Z-SSMNet), which adaptively integrates multi-dimensional (2D/2.5D/3D) convolutions to learn dense intra-slice information and sparse inter-slice information of the anisotropic bpMRI in a balanced manner. We also propose a self-supervised learning (SSL) technique that effectively captures both intra-slice and inter-slice semantic information using large-scale unlabeled data. Furthermore, we constrain the network to focus on the zonal anatomical regions to improve the detection and diagnosis capability of csPCa. We conducted extensive experiments on the PI-CAI (Prostate Imaging - Cancer AI) dataset comprising 10000+ multi-center and multi-scanner data. Our Z-SSMNet excelled in both lesion-level detection (AP score of 0.633) and patient-level diagnosis (AUROC score of 0.881), securing the top position in the Open Development Phase of the PI-CAI challenge and maintained strong performance, achieving an AP score of 0.690 and an AUROC score of 0.909, and securing the second-place ranking in the Closed Testing Phase. These findings underscore the potential of AI-driven systems for csPCa diagnosis and management.

Sarcopenia is associated with worse survival in non-small cell lung cancer (NSCLC), but less studied in association with toxicity. Here, we investigated the association between imaging-assessed sarcopenia with toxicity in patients with NSCLC. We analyzed a "chemoradiation" cohort (n = 318) of patients with NSCLC treated with chemoradiation, and an external validation "chemo-surgery" cohort (n = 108) who were treated with chemotherapy and surgery from 2002 to 2013 at a different institution. A deep-learning pipeline utilized pretreatment computed tomography scans to estimate SM area at the third lumbar vertebral level. Sarcopenia was defined by dichotomizing SM index, (SM adjusted for height and sex). Primary endpoint was NCI CTCAE v5.0 grade 3 to 5 (G3-5) toxicity within 21-days of first chemotherapy cycle. Multivariable analyses (MVA) of toxicity endpoints with sarcopenia and baseline characteristics were performed by logistic regression, and overall survival (OS) was analyzed using Cox regression. Sarcopenia was identified in 36% and 36% of patients in the chemoradiation and chemo-surgery cohorts, respectively. On MVA, sarcopenia was associated with worse G3-5 toxicity in chemoradiation (HR 2.00, P < .01) and chemo-surgery cohorts (HR 2.95, P = .02). In the chemoradiation cohort, worse OS was associated with G3-5 toxicity (HR 1.42, P = .02) but not sarcopenia on MVA. In chemo-surgery cohort, worse OS was associated with sarcopenia (HR 2.03, P = .02) but not G3-5 toxicity on MVA. Sarcopenia, assessed by an automated deep-learning system, was associated with worse toxicity and survival outcomes in patients with NSCLC. Sarcopenia can be utilized to tailor treatment decisions to optimize adverse events and survival.

Cerebrovascular diseases (CVD) can lead to stroke and dementia. Stroke is the second leading cause of death world wide and dementia incidence is increasing by the year. There are several markers of CVD that are visible on brain imaging, including: white matter hyperintensities (WMH), acute and chronic ischaemic stroke lesions (ISL), lacunes, enlarged perivascular spaces (PVS), acute and chronic haemorrhagic lesions, and cerebral microbleeds (CMB). Brain atrophy also occurs in CVD. These markers are important for patient management and intervention, since they indicate elevated risk of future stroke and dementia. We systematically reviewed automated systems designed to support radiologists reporting on these CVD imaging findings. We considered commercially available software and research publications which identify at least two CVD markers. In total, we included 29 commercial products and 13 research publications. Two distinct types of commercial support system were available: those which identify acute stroke lesions (haemorrhagic and ischaemic) from computed tomography (CT) scans, mainly for the purpose of patient triage; and those which measure WMH and atrophy regionally and longitudinally. In research, WMH and ISL were the markers most frequently analysed together, from magnetic resonance imaging (MRI) scans; lacunes and PVS were each targeted only twice and CMB only once. For stroke, commercially available systems largely support the emergency setting, whilst research systems consider also follow-up and routine scans. The systems to quantify WMH and atrophy are focused on neurodegenerative disease support, where these CVD markers are also of significance. There are currently no openly validated systems, commercially, or in research, performing a comprehensive joint analysis of all CVD markers (WMH, ISL, lacunes, PVS, haemorrhagic lesions, CMB, and atrophy).