Radiation Medicine and Protection

Sekkat H., Khallouqi A., Rhazouani O.E., Halimi A.

Reisdorf P., Gavrysh J., Ammann C., Fenski M., Kolbitsch C., Lange S., Hennemuth A., Schulz-Menger J., Hadler T.

Abstract Cardiovascular magnetic resonance imaging (CMR) offers state-of-the-art myocardial tissue differentiation. The CMR technique late gadolinium enhancement (LGE) currently provides the noninvasive gold standard for the detection of myocardial fibrosis. Typically, thresholding methods are used for fibrotic scar tissue quantification. A major challenge for standardized CMR assessment is large variations in the estimated scar for different methods. The aim was to improve quality assurance for LGE scar quantification, a multi-reader comparison tool “Lumos” was developed to support quality control for scar quantification methods. The thresholding methods and an exact rasterization approach were implemented, as well as a graphical user interface (GUI) with statistical and case-specific tabs. Twenty LGE cases were considered with half of them including artifacts and clinical results for eight scar quantification methods computed. Lumos was successfully implemented as a multi-level multi-reader comparison software, and differences between methods can be seen in the statistical results. Histograms visualize confounding effects of different methods. Connecting the statistical level with the case level allows for backtracking statistical differences to sources of differences in the threshold calculation. Being able to visualize the underlying groundwork for the different methods in the myocardial histogram gives the opportunity to identify causes for different thresholds. Lumos showed the differences in the clinical results between cases with artifacts and cases without artifacts. A video demonstration of Lumos is offered as supplementary material 1. Lumos allows for a multi-reader comparison for LGE scar quantification that offers insights into the origin of reader differences.

Paalvast O., Sevenster M., Hertgers O., de Bliek H., Wijn V., Buil V., Knoester J., Vosbergen S., Lamb H.

Abstract Despite the approval of over 200 artificial intelligence (AI) applications for radiology in the European Union, widespread adoption in clinical practice remains limited. Current assessments of AI applications often rely on post-hoc evaluations, lacking the granularity to capture real-time radiologist-AI interactions. The purpose of the study is to realise the Radiology AI lab for real-time, objective measurement of the impact of AI applications on radiologists’ workflows. We proposed the user-state sensing framework (USSF) to structure the sensing of radiologist-AI interactions in terms of personal, interactional, and contextual states. Guided by the USSF, a lab was established using three non-invasive biometric measurement techniques: eye-tracking, heart rate monitoring, and facial expression analysis. We conducted a pilot test with four radiologists of varying experience levels, who read ultra-low-dose (ULD) CT cases in (1) standard PACS and (2) manually annotated (to mimic AI) PACS workflows. Interpretation time, eye-tracking metrics, heart rate variability (HRV), and facial expressions were recorded and analysed. The Radiology AI lab was successfully realised as an initial physical iteration of the USSF at a tertiary referral centre. Radiologists participating in the pilot test read 32 ULDCT cases (mean age, 52 years ± 23 (SD); 17 male; 16 cases with abnormalities). Cases were read on average in 4.1 ± 2.2 min (standard PACS) and 3.9 ± 1.9 min (AI-annotated PACS), with no significant difference (p = 0.48). Three out of four radiologists showed significant shifts (p < 0.02) in eye-tracking metrics, including saccade duration, saccade quantity, fixation duration, fixation quantity, and pupil diameter, when using the AI-annotated workflow. These changes align with prior findings linking such metrics to increased competency and reduced cognitive load, suggesting a more efficient visual search strategy in AI-assisted interpretation. Although HRV metrics did not correlate with experience, when combined with facial expression analysis, they helped identify key moments during the pilot test. The Radiology AI lab was successfully realised, implementing personal, interactional, and contextual states of the user-state sensing framework, enabling objective analysis of radiologists’ workflows, and effectively capturing relevant biometrics. Future work will focus on expanding sensing of the contextual state of the user-state sensing framework, refining baseline determination, and continuing investigation of AI-enabled tools in radiology workflows.

Yan D., Li Q., Chuang Y., Lin C., Shieh J., Weng W., Tsui P.

Mun S.B., Choi S.T., Kim Y.J., Kim K.G., Lee W.S.

Manolakis D., Bizopoulos P., Lalas A., Votis K.

Abstract Ensuring strict medical data privacy standards while delivering efficient and accurate breast cancer segmentation is a critical challenge. This paper addresses this challenge by proposing a lightweight solution capable of running directly in the user’s browser, ensuring that medical data never leave the user’s computer. Our proposed solution consists of a two-stage model: the pre-trained nano YoloV5 variation handles the task of mass detection, while a lightweight neural network model of just 20k parameters and an inference time of 21 ms per image addresses the segmentation problem. This highly efficient model in terms of inference speed and memory consumption was created by combining well-known techniques, such as the SegNet architecture and depthwise separable convolutions. The detection model manages an mAP@50 equal to 50.3% on the CBIS-DDSM dataset and 68.2% on the INbreast dataset. Despite its size, our segmentation model produces high-performance levels on the CBIS-DDSM (81.0% IoU, 89.4% Dice) and INbreast (77.3% IoU, 87.0% Dice) dataset.

Ibrahim S., Selim S., Elattar M.

Niknejad Mazandarani F., Babyn P., Alirezaie J.

Ibrahim A.U., Engo G.M., Ame I., Nwekwo C.W., Al-Turjman F.

Chen H., Liu C., Cheng X., Jiang C., Wang Y.

Zhang T., Li R., Zhong Z., Zhang X., Liu T., Zhou G., Lv F.

Loizidou K., Skouroumouni G., Savvidou G., Constantinidou A., Vlachou E.O., Yiallourou A., Pitris C., Nikolaou C.

Abstract The objective is to predict a possible near-term occurrence of a breast mass after two consecutive screening rounds with normal mammograms. For the purposes of this study, conducted between 2020 and 2024, three consecutive rounds of mammograms were collected from 75 women, 46 to 79 years old. Successive screenings had an average interval of $$\sim$$ ∼ 2 years. In each case, two mammographic views of each breast were collected, resulting in a dataset with a total of 450 images (3 × 2 × 75). The most recent mammogram was considered the “future” screening round and provided the location of a biopsy-confirmed malignant mass, serving as the ground truth for the training. The two normal previous mammograms (“prior” and “current”) were processed and a new subtracted image was created for the prediction. Region segmentation and post-processing were, then, applied, along with image feature extraction and selection. The selected features were incorporated into several classifiers and by applying leave-one-patient-out and k-fold cross-validation per patient, the regions of interest were characterized as benign or possible future malignancy. Study participants included 75 women (mean age, 62.5 ± 7.2; median age, 62 years). Feature selection from benign and possible future malignancy areas revealed that 14 features provided the best classification. The most accurate classification performance was achieved using ensemble voting, with 98.8% accuracy, 93.6% sensitivity, 98.8% specificity, and 0.96 AUC. Given the success of this algorithm, its clinical application could enable earlier diagnosis and improve prognosis for patients identified as at risk.

Mehrnia S.S., Safahi Z., Mousavi A., Panahandeh F., Farmani A., Yuan R., Rahmim A., Salmanpour M.R.

Fayemiwo M., Gardiner B., Harkin J., McDaid L., Prakash P., Dennedy M.

Abstract Accurate segmentation of adrenal glands from CT images is essential for enhancing computer-aided diagnosis and surgical planning. However, the small size, irregular shape, and proximity to surrounding tissues make this task highly challenging. This study introduces a novel pipeline that significantly improves the segmentation of left and right adrenal glands by integrating advanced pre-processing techniques and a robust post-processing framework. Utilising a 2D UNet architecture with various backbones (VGG16, ResNet34, InceptionV3), the pipeline leverages test-time augmentation (TTA) and targeted removal of unconnected regions to enhance accuracy and robustness. Our results demonstrate a substantial improvement, with a 38% increase in the Dice similarity coefficient for the left adrenal gland and an 11% increase for the right adrenal gland on the AMOS dataset, achieved by the InceptionV3 model. Additionally, the pipeline significantly reduces false positives, underscoring its potential for clinical applications and its superiority over existing methods. These advancements make our approach a crucial contribution to the field of medical image segmentation.

Han B., Yang Q., Tao X., Wu M., Yang L., Deng W., Cui W., Luo D., Wan Q., Liu Z., Zhang N.