BACKGROUND:Ghana's imaging services face rising demand, uneven digital infrastructure, and limited access to advanced modalities. Artificial intelligence (AI) could improve diagnostic accuracy, workflow efficiency, and access, but real-world adoption is early. OBJECTIVE:Assess Ghana's readiness to adopt AI in medical imaging, identify pathways and barriers, and propose a phased, context-specific roadmap. METHODS:A review of peer-reviewed and grey literature (2012-March 2025) using PubMed, IEEE Xplore, Scopus, Google Scholar, and Ghanaian institutional documents. The review emphasizes imaging AI, LMIC experiences, and Ghana-specific evidence on infrastructure, policy, and pilots, distinguishing Ghana-based findings from international evidence extrapolated to Ghana. RESULTS:Major gaps include digital infrastructure (limited PACS, variable DR/CR adoption, uneven connectivity), financing (license and maintenance costs), governance (SaMD pathways exist but AI-specific provisions are evolving; operational data protection needs strengthening), and workforce (limited AI literacy; urban - rural disparities). Ghana-relevant touchpoints include MinoHealth.AI chest radiography evaluations, the national imaging equipment inventory, Ghana Health Service digital health strategy (2023-2027), and FDA SaMD guidance. A phased roadmap is proposed: establish PACS and connectivity; implement AI governance and data stewardship; run targeted pilots in CXR triage, low-dose CT, and MRI acceleration; scale via public-private partnerships and pooled procurement; and sustain workforce development with human-in-the-loop oversight. CONCLUSIONS:AI can improve equity and efficiency in imaging in Ghana if adoption builds on strong digital foundations, robust governance, local validation, and clinician-led implementation. Priorities include PACS deployment, AI-specific regulatory strengthening, ethical data governance, and capacity building to support safe, equitable, and sustainable use.
INTRODUCTION:Osteoarthritis (OA) is a degenerative joint disease characterized by cartilage loss, synovial fluid imbalance, and bone structural changes, leading to reduced mobility. Most clinical studies use MRI-derived cartilage characteristics to assess OA progression. To support timely treatment decisions and minimize human error, an automated computer aided system is needed for prediction of OA in the progressive stages. METHODS:To build the automatic system for classifying progression phases of OA, we present a novel hybrid framework that incorporates cartilage characteristics of longitudinal knee MRI images from OAI dataset. The framework consists of two key phases: (i) initially, cartilage regions were segmented using an attention-enhanced position-aware encoder-decoder network. Multiple attention mechanisms were investigated at different network locations to identify the optimal segmentation strategy, and (ii) Morphological shape features were extracted from the segmented from the segmented MRI images. Statistical analysis was performed to select the most discriminative features, which were then used to classify OA progression stages corresponding to 18-month and 30-month follow-up examinations using machine learning classifiers. RESULTS:The results demonstrated that the proposed framework effectively classified OA progression at both follow-up periods. Among all classifiers, Random Forest achieved the best performance with an average F-measure of 84.13%, specificity of 87.65%, sensitivity of 96.15%, and accuracy of 86.67% across all folds. Overall, the approach attained average accuracy and F-measure of 76.50% and 71.84% respectively. CONCLUSIONS:The effectiveness of this framework suggests its potential as an automatic decision-support system to assist clinicians in making accurate and consistent OA diagnosis.
INTRODUCTION:Volumetric modulated arc therapy (VMAT) for non-small cell lung cancer (NSCLC) may result in radiation-induced haematological toxicities. To estimate the dose received by circulating blood particles, we implemented a four-dimensional dose delivery algorithm into a stochastic simulation of blood dynamics. METHODS:The treatment data of NSCLC patients who received VMAT between 2018 and 2024 were considered. The prescribed dose was 60 Gy/55 Gy/50 Gy delivered in 30/20/25 fractions. To estimate the blood dose, the doses at each control point calculated with the BeamSplitter were used in the Haematological Dose software (HEDOS). The blood dose metrics D98%/mean dose/D2% were calculated and compared with the four-dimensional algorithm (variable dose rates) and the original algorithm (constant dose rates). The main organs contributing to the blood dose were reported. RESULTS:The cohort consisted of 115 treatments of 98 patients. As compared with constant dose rates, the population blood D98% decreased (median ΔD98% = -0.1 Gy, p-value < 0.001) with variable dose rates, whilst blood D2% increased (median ΔD2% = 0.1 Gy, p-value < 0.001). The mean blood doses were similar in both calculations (median mean blood dose = 5.4 Gy, p-value = 0.3). At the population level, the largest change in blood D2% was 0.38 Gy. The main organs contributing to the blood dose were lungs (25%), brachiocephalic veins (14%), and aorta (12%). CONCLUSIONS:VMAT for NSCLC induced a larger blood D2% and a smaller blood D98% as compared with constant delivery.
PURPOSE:In breast radiotherapy, delivery of manually-calculated electron boosts limits retrospective dose-response analyses as dose distribution is unavailable. This work evaluates the feasibility of reconstructing dose distributions from manually planned electron boosts in breast-conserving radiotherapy. METHODS:Only 72 out of 198 breast cancer patients had complete stored dose distributions from sequential electron boosts in the REQUITE study. Arbitrary data from 70/72 patients were used to develop and validate dose reconstruction method. Twenty patients were used to determine optimal parameters for Monte-Carlo-based (MC) electron dose reconstruction on RayStation (v.11B-R), considering CT-calibration curve, MC-history number, andcalculation grid resolution. Remaining 50 patients were used to quantify dose reconstruction accuracy. The similarity between reconstructed and stored dose was evaluated using 3D-gamma index and dosimetric parameters extracted from breast and tumour bed contours. Dose difference location was evaluated using dose-location histogram. RESULTS:Calculation grid resolution significantly impacted electron dose distribution (p < 0.01), where the finest grid (0.15 cm) showed highest similarity to stored doses. CT-calibration curve and MC-history number had a negligible influence on dose reconstruction. Dosimetric difference between reconstructed and stored doses was < 1 Gy for breast and tumour bed. Reconstructed dose was achieved > 90% gamma passing rate in the validation set. However, around 2.5 Gy dose differences were observed at the skin and tissue interface regions. CONCLUSIONS:Retrospective electron boost dose reconstruction is feasible with acceptable accuracy, and could increase data completeness in large cohort studies. Caution is advised when assessing dose near tissue interface and further validation is needed outside the REQUITE dataset.
Background and purpose Pancreatic SBRT planning is challenging due to the proximity of highly sensitive organs at risk. Since automated planning is currently not available for the CyberKnife (CK) system, the optimization of high-quality treatment plans relies heavily on expert planners; however, even in this setting, the generation of suboptimal plans remains possible. The aim of this work was to develop a tool to guide plan optimization based on simple geometric parameters and institutional planning experience. Materials and methods As a first step, the predictive value of the Expansion-Intersection Volume (EIV), defined as the intersection volume between the PTV expanded by 5 mm and the duodenum, stomach, and bowel, was evaluated together with GTV and PTV volumes. The investigated outputs included PTVV40Gy [%], GTVV47.5 Gy [%], GTVV50Gy [%], monitor units (MUs), delivery time, and a plan complexity score. A machine learning-based tool was then developed to predict these outputs for new patients using information extracted from 41 previously optimized plans, while also identifying the most similar historical cases. Results A moderate-to-strong significant correlation was observed between the three input parameters and the three dosimetric outputs. A similar relationship was found for the remaining plan efficiency and complexity metrics. The tool was tested on 10 new patient plans, and predicted values were compared with the corresponding software-guided planning results. Conclusions A simple knowledge-based planning-support tool was developed. Using only three input parameters (EIV, PTV volume, and GTV volume), the tool provides an estimate of achievable target coverage, plan efficiency, and complexity for CK pancreatic SBRT. In addition, it identifies similar historical cases, providing a useful starting point for plan optimization and helping to reduce the risk of suboptimal planning outcomes.
BACKGROUND:Intravoxel incoherent motion (IVIM) modelling of diffusion-weighted magnetic resonance imaging (DW-MRI) quantifies tissue diffusion (Dt), capillary microcirculation (Dp) and perfusion fraction (Fp). However, accurate IVIM modelling using the few b-values acquired in clinical settings is challenging. This study mainly investigated the performance of four algorithms for accurate IVIM parameter modelling with few b-values before exploring the IVIM parameters' biomarker potential for head and neck cancer (HNC) radiotherapy response. METHODS:Synthetic IVIM signals with 4, 5 and 11 b-values were generated. DW-MRI was acquired with 11 b-values for 10 HNC patients. Self-supervised (DNNSSL) and supervised (DNNSL) deep neural networks, and least square (LSQ) and segmented (SEG) fitting algorithms estimated IVIM parameters for both synthetic and patient imaging data using 4, 5 and 11 b-values. The algorithm accuracy was evaluated in silico and in vivo by calculating the error between the estimated and ground truth parameter values. A Kaplan-Meier analysis evaluated the associations between progression free survival (PFS) of 20 HNC patients and longitudinal changes in IVIM parameters estimated with 4 b-values. RESULTS:LSQ and DNNSL were the least and most accurate IVIM modelling algorithms, respectively. DNNSL, however, was biased towards the mean which limits the parameters' biomarker potential. None of the algorithms, however, estimated Dp accurately with few b-values. Preliminary survival analysis suggested that an increase in Fp during radiotherapy reduced the PFS of HNC patients. CONCLUSIONS:SEG and DNNSSL were the preferred algorithms for IVIM modelling few b-values, although reliable estimation was limited to the Dt and Fp parameters.
BACKGROUND:The study assessed the role of CT-based breast tissue composition in predicting toxicities after moderately hypofractionated whole-breast postoperative Radiotherapy. MATERIALS AND METHODS:A mono-institutional cohort of 1127 patients (2009-2017) treated with 40 Gy/15 fractions was analyzed. Endpoints included acute RTOG toxicities (ACUG2 + ), 6-month edema/hyperpigmentation (EDMG2 + ), 3-year Fibrosis-Atrophy-Telangiectasia-Pain (FATP) and Liponecrosis (LIPNEC). HU-histograms (-200 to + 70 HU) were extracted from planning CTs using custom Python code, deriving mean/median HU, SD, kurtosis, skewness, percentiles, and fat/fibroglandular volumetric fractions. The population was split into training/validation (800/327). Univariate and Multivariate Logistic Regression and a bootstrap-based machine learning approach were performed, including CTV volume as a potential predictor. RESULTS:ACUG2+/EDMG2+/FATP/LIPNEC rates were 14.5%/4.5%/4.3%/11% respectively. CTV volume was significantly associated with all endpoints. No HU-histogram parameter increased the predictive power of CTV volume for ACUG2+/EDMG2+/FATP (AUC_val = 0.56/0.65/0.68). For LIPNEC, a model combining HU-Mean and the count of the maximum HU value of the histogram was the most predictive (AUC_val = 0.58). Subgroup analyses stratified by CTV volume revealed that a larger adipose fraction may exert a protective effect on acute skin reactions and edema in patients with smaller breasts. CONCLUSIONS:HU-histogram parameters show limited association with toxicities compared to CTV volume. However, subgroup analyses suggest a volume-dependent contribution of breast tissue composition, with adipose tissue playing a protective role for ACUG2 + and EDMG2 + in smaller-volume strata.
PURPOSE:To analyze dosimetric data and establish local typical dose values (LTDVs) for otolaryngology scans performed on cone-beam CT (CBCT) MATERIALS AND METHODS: Dosimetric indicators were retrospectively collected for otorhinolaryngology examinations performed on a CBCT device. LTDVs were determined for each examination type using weighted computed tomography air kerma index (Cw), air kerma-length product (PKL), kerma-area product (PKA) and acquisition length. Cw, PKL and length were compared according to the acquisition protocols. CBCT LTDVs were compared with those obtained using multi-slice CT (MSCT) for inner ear and sinus examinations. RESULTS:Inner ear examinations showed significantly higher Cw, PKL and PKA values than other examination types, with mean increases of 206.7 ± 15.8%, 29.0 ± 3.7% and 17.0 ± 2.3%, respectively (p < 0.001 and 0.50 ≤ ES ≤ 0.87). Salivary gland and sinus examinations showed similar values (0.503 ≤ p ≤ 0.749), while facial massif examinations showed significantly lower values -8.4 ± 0.0%, -4.8 ± 0.0% and -3.2 ± 0.6%, respectively (p < 0.041 and ES ≤ 0.20). The Cw decreased as the FOV increased and the opposite was found for PKL and acquisition length. CBCT's LTDVs were lower than with MSCT for the inner ear by 43.9 ± 1.7%. For sinuses, CBCT's LTDVs were 30% lower than the national diagnostic reference levels (NDRL) for Cw and 39.5% lower for PKL but 43.3 ± 1.9% higher than the LTDVs for MSCT. CONCLUSION:In CBCT-based ENT imaging, inner ear examinations involve the highest radiation doses, although these remain lower than with MSCT. Sinus doses are lower than MSCT NDRLs but higher than LTDVs.
BACKGROUND:An increasing number of pregnant patients undergo radiological examinations or procedures each year, raising concerns about fetal radiation exposure. Several methods for fetal dose (FD) estimation are available. Within the EU-funded SONORA project, this study aims to identify, evaluate, and synthesize existing FD estimation methods used in diagnostic and interventional radiology (DIR). The objective is to compare evidence from the literature with real-world clinical practices across Europe, identifying discrepancies and areas requiring harmonization. MATERIALS AND METHODS:A systematic review of PubMed and Embase databases (2000-2025) was conducted following PRISMA guidelines, including 75 original studies. In parallel, a survey was distributed across European countries to investigate current FD estimation practices, tools, reported dose ranges, and optimization strategies. Fifty-two valid responses, mainly from Medical Physicists, were analyzed. RESULTS:Literature and survey data identified three main FD estimation approaches: Monte Carlo simulations, physical phantom measurements, and software tools, the latter being most used in clinical practice. Reported FD values varied by imaging modality: 10-60 mGy for abdominal or pelvic CT, <1 mGy for chest CT and conventional radiography, and a wide range (0.01->100 mGy) for interventional procedures. While strong agreement was observed on technical optimization strategies-such as collimation, tube current modulation, and reduction of CT acquisition phases-substantial variability persists in the choice of software tools. CONCLUSIONS:Despite the growing use of ionizing radiation during pregnancy, FD estimation methods remain heterogeneous. Harmonization of international guidelines and development of standardized tools incorporating maternal biometric data are urgently needed to improve risk assessment and ensure safer management of pregnant patients.
PURPOSE:The performance of the Monte Carlo code PENHAN is benchmarked for the simulation of the microdosimetric spectra of low-energy protons (≤ 14 MeV) using 3D-cylindrical silicon detectors. METHODS:Simulations were validated against experimental spectra obtained from a proton beam cyclotron at the Centro Nacional de Aceleradores (CNA, Seville) and compared with results from the TOPAS (Geant4) code, using the same methodology reported in the paper that presented the reference experimental data. RESULTS:The results showed that PENHAN reproduced the shape and position of the energy spectra, matching the maximum peak of the energy distribution to within 6% relative error in all cases. Absolute relative errors in the frequency-mean lineal energy remained below 8% across all proton energies (6-14 MeV). We also emphasize the importance of accurate modeling of the electronic stopping power, Se, for protons: variations in Se for tungsten alone led to differences of up to 14% in frequency-mean lineal energy. CONCLUSION:These findings underscore the potential of PENHAN as a simulation tool in proton therapy and microdosimetry, and highlight the necessity of validated stopping power data for reliable simulations.
INTRODUCTION:Optimizing uptake assessment schedules is critical for minimizing uncertainties in internal dosimetry. We performed a simulation study based on patient-specific biokinetics from iodine-131 (131I)-sodium iodide (NaI) hyperthyroidism treatments to evaluate the impact of measurement precision and sampling timing on the bias and precision of the time-integrated activity (TIA) estimates. METHODS:Patient biokinetics were simulated using a two-compartment model, with parameters derived from the distribution of a cohort undergoing planning dosimetry. Uptake measurements were then simulated with variable relative noise for each simulated patient. The bias and precision were calculated as the mean and standard deviation, respectively, of the difference between the fitted and true TIA values. RESULTS:Graphs illustrate the estimated TIA bias and precision as functions of measurement precision for various uptake assessment schedules; values for 5% and 10% measurement precision are provided in tables. CONCLUSION:For the uptake assessment schedules considered, the TIA bias is negative and for schemes with more than one uptake measurements it depends on measurement precision with increasing magnitude. The precision and bias for the (5, 24, 168)h and (24, 168)h post-administration uptake assessment schemes are comparable. The last uptake assessment should not be performed earlier than 6 days post-administration to keep the TIA precision below 15% for a 10% measurement precision. The use of error propagation and the Hessian matrix for curve fit parameter uncertainty estimation leads to an overestimation of the true curve integral uncertainty for large uptake assessment uncertainties.
PURPOSE:The increasing integration of Artificial Intelligence (AI) into clinical workflows for medical imaging and radiotherapy presents new opportunities and challenges for the radiation protection of patients, staff, and the public. This perspective from the International Commission on Radiological Protection (ICRP) Committee 3 Working Party on AI examines how current and emerging AI applications support the core principles of justification and optimisation across diagnostic and interventional radiology, nuclear medicine, and radiotherapy, and identifies priorities for their safe clinical implementation. MAIN FINDINGS:AI applications with the greatest current clinical maturity include clinical decision support for referral appropriateness, image reconstruction, protocol optimisation, automated contouring, treatment planning, adaptive radiotherapy, and AI-enabled quality assurance. Other applications, including patient-specific dosimetry, occupational dose prediction, synthetic imaging, and predictive safety analytics, show considerable promise but remain at earlier stages of validation. Significant challenges accompany these advances: data biases and limited generalisability may undermine performance across diverse settings; the "black box" nature of many models complicates clinical accountability; and robust validation, continuous quality assurance, and harmonised regulatory oversight remain essential. Dedicated training in AI literacy for healthcare professionals is critical for safe deployment. CONCLUSION:AI has great potential to improve medical radiation protection but its safe use requires careful management of associated risks. By identifying areas of established clinical adoption, emerging applications, and common implementation priorities, this perspective provides a framework to support future ICRP recommendations on the safe integration of AI into medical radiation protection.
BACKGROUND:Ischemic stroke results from the occlusion of a cerebral artery and is a leading cause of mortality and disability worldwide. Multimodal computed tomography (CT), including CT perfusion (CTP) and CT angiography, is crucial to acute stroke evaluation but involves higher radiation exposure than non-contrast CT due to repeated volumetric imaging. Reducing CTP radiation dose without losing image quality remains important challenge. This study proposes a machine-learning-based denoising autoencoder (DAE) to reduce noise introduced by dose reduction while preserving the quality of CTP images and perfusion parameter maps. METHOD:CTP images from 48 acute ischemic stroke patients from the PRove-IT trial were used. Low-dose conditions were simulated by adding Gaussian and Poisson noise at varying strengths, with Poisson noise applied in the sinogram domain and Gaussian noise in the image domain. The DAE was trained using paired noisy and original images. Performance was evaluated by assessing structural similarity of CTP source images and perfusion maps, as well as clinical accuracy based on infarct core volumes derived from cerebral blood flow maps. RESULTS:The DAE restored strong structural similarity in CTP source images at dose reductions up to 90% (SSIM 0.81, PSNR 43 dB). Perfusion maps showed slightly lower similarity. Clinically, denoising substantially improved the accuracy of CBF-derived infarct core volumes, reducing mean absolute error from 10-30 mL in noisy images to approximately 4-16 mL and restoring high agreement with reference volumes (R2 > 0.85). CONCLUSIONS:These findings demonstrate that substantial simulated radiation dose reductions can be compensated by the DAE while preserving clinically meaningful perfusion-derived biomarkers.
PURPOSE:This study evaluates the feasibility, dosimetric optimization, and delivery efficiency of c_arm linac-based and TomoTherapy-based Lattice Radiotherapy (LRT) across head and neck, thoracic, and abdominal tumors. MATERIALS AND METHODS:72 patients meeting inclusion criteria (GTV > 480 cc, at least eight lattice vertices) were planned with both VMAT and Helical techniques. Each high-dose lattice received 60 Gy in five fractions. Identical lattice geometries and concentric valley dose rings were applied for both systems. VMAT plans utilized four partial arcs, while helical plans employed helical delivery with a 2.5 cm field width to assess the impact on dose modulation. Key dosimetric indices (GTV D95%, Dmean, D50%, Homogeneity Index, Peak-to-Valley Dose Ratio, and spatial dose modulation metrics) were calculated. An anthropomorphic phantom with OSLDs and ArcCHECK was used for independent validation. RESULTS:Both platforms achieved the intended spatially fractionated dose distribution with adequate vertex coverage. c_arm plans demonstrated significantly higher peak doses (D2%) and PVDR, indicating stronger peak-valley modulation. ring_gantry plans showed higher D95% and D98%, reflecting increased low-dose GTV coverage and smoother intratumoral dose redistribution, whereas c_arm plans showed higher PVDR, indicating sharper peak-valley modulation. Beam-on time was substantially shorter for c_arm (3-5 min) compared to ring_gantry (20-31 min, p < 0.001). CONCLUSION:Both platforms generated acceptable LRT dose distributions with distinct platform-level dosimetric characteristics. c_arm plans produced sharper peak-valley modulation and shorter beam-on times, whereas ring_gantry plans showed smoother intratumoral dose redistribution, modest reductions in selected OAR dose metrics, and slightly higher gamma agreement under the evaluated QA criteria.
AIM:To investigate and quantify the influence of the pulsed nature of clinical irradiation beams on the response of a specific type of active personal dosimeter (APD), and to evaluate its impact on individual occupational dose measurements of interventional radiologists and cardiologists. METHODS:APD responses were compared against passive dosimeters in a controlled experimental setting. Deviations outside the devices' specifications were addressed using a two-step correction model. A correction method based on the non-paralyzable dead-time model was developed using the instantaneous dose rate per pulse extracted from the Radiation Dose Structured Report (RDSR) at the irradiation-event level. After determining the appropriate correction factor, the method was applied to one year of clinical data comprising approximately 250,000 irradiation events. RESULTS:Among 24 occupational APDs evaluated, 14 required corrections, with the largest addition to the cumulative dose being 3.1 ± 0.1%. Five dosimeters mounted on C-arms showed that 6.61% of recorded events required correction, with a maximum correction of 5.52 ± 0.02%. Overall, the impact of pulsed radiation on recorded occupational doses was limited. CONCLUSIONS:Although the overall influence of pulsed radiation beams on APD-recorded doses was modest, verification of potential count-loss effects remains important to ensure the accuracy of cumulative dose measurements. With appropriate correction measures, APDs can be reliably used to support optimization of radiological protection for interventional specialists.
BACKGROUND AND AIM:Microbeam Radiotherapy (MRT) is a spatially fractionated radiotherapy modality consisting of high and low dose regions called peaks and valleys. It has been demonstrated that the prohibitively long computational time required for Monte Carlo dose calculation can be addressed effectively via the application of 3D U-Nets, but these networks can be inaccurate, especially for the valley dose. Recursive Convolutional Layers (RCL) have been shown to improve the performance of neural networks, however have never been applied to MRT dose prediction. This work applies RCLs to the problem of predicting Geant4 simulated MRT dose in rats, with the aim of increasing the dose prediction accuracy. METHODS:An existing 3D U-Net architecture for MRT is modified to contain RCLs. The effect of RCLs with five recursions is investigated for every location of the 3D U-Net, for peak and valley dose. To investigate the effect of RCLs increasing the model's virtual depth, parts of the 3D U-Net are replaced with five sequential blocks of the same type. RESULTS:The RCLs improved the model accuracy for both peak and valley dose. The valley dose prediction had the greatest improvement, with the greatest increase from 70.6% phantom voxels agreeing within 3% of the Geant4 dose to 84.9% when using the RCLs. The five sequential blocks performed worse in predicting valley dose in bone and at tissue boundaries. It is further found that it is best to place RCLs immediately after the input and before the output within the same model.
PURPOSE:This study aimed to evaluate the prognostic value of FDG-PET/CT-based radiomic signatures of bone lesions for predicting overall survival (OS) in metastatic melanoma patients treated with immune checkpoint inhibition (ICI). MATERIAL AND METHODS:Thirty-six consecutive metastatic melanoma patients treated with ICI at the University Hospital Zurich were included in this study. All patients had at least one bone metastasis at baseline, and PET/CT imaging data were available. The clinical endpoint was defined as patient OS. Radiomic features were extracted from each lesion and were correlated with patient OS. For patients with multiple bone lesions, radiomic data were pooled using three methods: arithmetic mean, mean weighted by lesion volume and selection of the largest lesion. Multivariate models were internally validated with nested cross-validation, and the area under the receiver operating characteristic curve was used as the performance metric. RESULTS:The median OS was 5.1 years. Among the models predicting 3-year OS post-treatment, the highest performance was demonstrated by the combined PET/CT model integrating features from all aggregation methods, achieving an AUC of 0.76 (95% CI: 0.57-0.95). For 5-year OS prediction, the best performance was also observed for the combined PET/CT model, with an AUC of 0.84 (95% CI: 0.71-0.98), using radiomic features extracted from the largest lesions. CONCLUSIONS:FDG-PET/CT-based radiomic signatures of bone lesions in metastatic melanoma patients show potential for predicting overall survival. These findings suggest that radiomics could serve as a tool for guiding treatment decisions in patients with bone metastases undergoing immune checkpoint inhibition.
PURPOSE:To evaluate the dosimetric benefit of deep inspiration breath-hold (DIBH) using the Active Breathing Coordinator (ABC) for left-sided breast cancer radiotherapy and to identify patients most likely to benefit based on free-breathing (FB) anatomical and clinical features. METHODS:Thirty-one postmastectomy patients underwent FB and ABC CT scans for volumetric modulated arc therapy planning. Thirteen anatomical features and clinical variables were extracted from FB images. Correlation analyses were performed to assess associations between FB-derived features and dose changes in organs at risk. Logistic regression (LR) and gradient boosting (GB) models were developed to predict two endpoints: a reduction in mean heart dose (MHD) >20% and a reduction in mean left lung dose (MLD) >10% with ABC compared with FB. SHapley Additive exPlanations (SHAP) analysis was used to interpret the GB models. RESULTS:ABC significantly reduced cardiac and pulmonary doses. Cardiopulmonary volume ratio (CVR) was strongly associated with mean MHD reduction, and patients with CVR >0.21 were more likely to achieve cardiac sparing with ABC. For predicting an MHD reduction >20%, the 5-feature GB model achieved an AUC of 0.93. For predicting an MLD reduction >10%, the 6-feature GB model achieved an AUC of 0.92. SHAP analysis confirmed the key role of CVR in predicting cardiac benefit and highlighted right lung volume as an indicator of inspiratory expansion potential. CONCLUSIONS:FB-based anatomical features can preliminarily identify patients likely to benefit from ABC. Interpretable ML models further improve patient selection and may support the use of ABC in breast cancer radiotherapy.
