Lung nodule segmentation on computed tomography (CT) is at the same time one of the most common and laborious tasks in oncological radiology. Fortunately, artificial intelligence agents have been showing promising results in streamlining the process. We study some of the challenges of training an AI model for lung nodule segmentation, including the degradation of performance due to distribution shift, privacy concerns and limited bandwidth for cloud data transmission. The article explores different federated learning strategies, over a pool of 1506 CT studies collected from four hospitals. The results show that federated learning models reach near standard classical training DICE score performance (i.e., 87.24% vs. 88.96%), and even surpass it in a privacy-centered context (i.e., 87.24% vs. 84.78%). Additionally, active learning was proven to increase the new model’s DICE score by 1.76% over the random sampling strategy. The article adds to the growing body of research exploring the use of federated learning in healthcare and demonstrates its potential for improving lung nodule segmentation on CT.

Domain shift poses significant problems to computer-aided diagnostic (CAD) systems when deployed in clinical scenarios. There’s still no definite fix nor an in-depth understanding of the exact factors driving domain shifts in medical X-rays. Here, we conduct an exploratory study on three covariate shift factors in X-ray classification by controlling for different variables. This is possible by leveraging a homogenously-relabelled mix of public and private X-ray data spanning 23 medical institutions over four continents and 17 classes of pathologies. We show that the acquisition parameter, device manufacturer and geographical shifts degrade out-of-distribution (OOD) F1 by 6%, 3.2% and 3.3%, respectively. Pneumothorax was found to be the most impaired pathology, suffering a mean F1 generalisation gap of 13.3%, despite being one of the most clinically-consequential radiological findings. To this end, we introduced LISA-topK, a multi-label adaptation of Learning Invariant Predictors with Selective Augmentation (LISA), that we showed to narrow down the OOD gap, surpassing other methods consistently. These pragmatic results shed light on some of the elements of OOD generalisation in X-ray classification, which are essential to researching, understanding and deploying CAD systems. Code is available at https://github.com/RayscapeAI/LISA-topK

Medical imaging plays a critical role in patient diagnosis, with CT scans being an essential modality that provides a fine-grained, 3-dimensional insight into patients’ bodies. The large amount of information generated by CT scans can lead to long interpretation times and increased errors, particularly when compounded by radiologist fatigue. One solution is the use of Computer Aided Diagnostic (CAD) tools, which studies have shown can improve accuracy and reduce diagnosis time when used by radiologists. However, CAD tools can be computationally and memory intensive with long inference times, particularly when low-end computers or edge devices are used. This article explores multiple optimization techniques using a 3D nn-UNet lung lobe segmentation architecture in CT scans and reduces the inference time of the segmentation models by 64.9% (without performance degradation) to 95.9% (with significant performance degradation). These optimizations can reduce the overall diagnostic time for radiologists, and enable the segmentation models to run on CPU-only machines and on single board computers, which are readily available in both developed and developing countries.

Chest computed tomography (CT) has played a valuable, distinct role in the screening, diagnosis, and follow-up of COVID-19 patients. The quantification of COVID-19 pneumonia on CT has proven to be an important predictor of the treatment course and outcome of the patient although it remains heavily reliant on the radiologist's subjective perceptions. Here, we show that with the adoption of CT for COVID-19 management, a new type of psychophysical bias has emerged in radiology. A preliminary survey of 40 radiologists and a retrospective analysis of CT data from 109 patients from two hospitals revealed that radiologists overestimated the percentage of lung involvement by 10.23 ± 4.65% and 15.8 ± 6.6%, respectively. In the subsequent randomised controlled trial, artificial intelligence (AI) decision support reduced the absolute overestimation error (P < 0.001) from 9.5% ± 6.6 (No-AI analysis arm, n = 38) to 1.0% ± 5.2 (AI analysis arm, n = 38). These results indicate a human perception bias in radiology that has clinically meaningful effects on the quantitative analysis of COVID-19 on CT. The objectivity of AI was shown to be a valuable complement in mitigating the radiologist's subjectivity, reducing the overestimation tenfold.Trial registration: https://Clinicaltrial.gov . Identifier: NCT05282056, Date of registration: 01/02/2022.

Lung nodule detection remains one of the most common and painstaking tasks in radiology. Efforts to aid overworked radiologists are made using artificial intelligence (AI), although computed tomography (CT) makes it a hard computational task in practical scenarios. This study analyses the translation of a weight-averaging ensemble technique, from natural image classification to small object detection on CT. A dataset of 1050 patients is used to fine-tune models under diverse configurations to compare different types of ensembles. The model soup boosts the FROC score from 0.872 to 0.886, with no computational downsides. Next, two radiologists test their detection performance with and without the ensemble assistance on 20 CT studies. The AI improves the physicians’ mean sensitivity from 91.2 ± 5.1% to 94.2 ± 4.6%, while preserving a non-inferior specificity (P < 0.001). These results further pave the way to translating general computer vision advancements into the medical domain, supporting AI’s place in the physician’s toolbox.

Classification and object detection are computer vision tasks with successful, clinical applications in medical imaging. Yet, the increased effort required of expert readers in order to annotate bounding boxes on medical images has yet to be quantitatively justified in terms of added value to identifying pathologies. In this study, we show preliminary results on the classification and localization of 17 most common chest pathologies on a private dataset of 15,000 radiographs from two Romanian public hospitals. Next, we quantitatively compare the extra added value of the bounding box information at training time, concerning classification performance improvements. Two types of architectures were trained on publicly available and private data, classification architectures (i.e., InceptionNet V3), tasked with identifying pathologies in chest radiographies and object detection architectures (i.e., Faster R-CNN), tasked with localizing the regions of interest on the image. Both achieved high classification performance (i.e., 90.52 and 88.94 mean AUROC, respectively). The object detector, however, reached superior classification performance, thus proving the additional bounding box information available at training time, benefits patient-level pathology identification as well.

Studies have shown that radiologists working together with Computer Aided Diagnostic software have increased accuracy. Automated screening software can be used to prioritize X-Rays coming in for diagnosis. We developed a suite of machine learning algorithms that aim to improve radiologist performance. It provides suggested diagnostics, a heatmap showing pathological areas and a bone subtracted version of the image which helps radiologists to identify fractures. We test different configurations for our diagnosis model, training it on both normal and enhanced images, using one or two branches. Our experiments show that adding enhanced inputs (lung segmented and bone subtracted versions of the input) increases the performance of our algorithm, which in turn increases the performance of the radiologist user. This shows that preprocessing the images before input increases model performance. More research is needed to find other preprocessing techniques, to refine existing ones, and to determine the optimal number and type of input X-Rays.

The decreasing costs and increasing efficiency of medical imaging equipment has made radiology a central tool in diagnosing diseases all around the world. This has created an opportunity to develop efficient computer vision algorithms with a large potential impact. Two segmentation models based on the U-Net architecture were trained on publicly available and private datasets, obtained from collaborating hospitals. They achieved good performance (0.989 and 0.983 using the DICE metric), that generalizes to new data distributions, on the task of automatically calculating the cardiothoracic ratio from frontal radiographies. When testing on previously unseen datasets with new data distributions to measure their ability to generalize we obtain average DICE scores of 0.968 and 0.963 respectively. This ratio can be used as a clear indicator of the abnormal enlargement of the heart, and can be linked with other chest pathologies. The algorithm was tested in collaboration with a team of 3 radiologists, in a real world hospital setting. On average, the time they spent analyzing the heart was decreased, from 22.53s to 5.14s, and their F1 score on the task of cardiomegaly detection was increased from 0.845 to 0.851. Localized energybased normalization was applied and an increase in generalization ability of the model was observed, which was tested using 5 fold cross validation on all of the datasets. We propose studying how lung and heart segmentation can be used to detect structural abnormalities.