Unsupervised landmark detection and classification of lung infection using transporter neural networks.

Supervised deep learning techniques have been very popular in medical imaging for various tasks of classification, segmentation, and object detection. However, they require a large number of labelled data which is expensive and requires many hours of careful annotation by experts. In this paper, an unsupervised transporter neural network framework with an attention mechanism is proposed to automatically identify relevant landmarks with applications in lung ultrasound (LUS) imaging. The proposed framework identifies key points that provide a concise geometric representation highlighting regions with high structural variation in the LUS videos. In order for the landmarks to be clinically relevant, we have employed acoustic propagation physics driven feature maps and angle-controlled Radon Transformed frames at the input instead of directly employing the gray scale LUS frames. Once the landmarks are identified, the presence of these landmarks can be employed for classification of the given frame into various classes of severity of infection in lung. The proposed framework has been trained on 130 LUS videos and validated on 100 LUS videos acquired from multiple centres at Spain and India. Frames were independently assessed by experts to identify clinically relevant features such as A-lines, B-lines, and pleura in LUS videos. The key points detected showed high sensitivity of 99% in detecting the image landmarks identified by experts. Also, on employing for classification of the given lung image into normal and abnormal classes, the proposed approach, even with no prior training, achieved an average accuracy of 97% and an average F1-score of 95% respectively on the task of co-classification with 3-fold cross-validation.

Interobserver Variability of Hip Dysplasia Indices on Sweep Ultrasound for Novices, Experts, and Artificial Intelligence.

BACKGROUND: Ultrasound for developmental dysplasia of the hip (DDH) is challenging for nonexperts to perform and interpret. Recording "sweep" images allows more complete hip assessment, suitable for automation by artificial intelligence (AI), but reliability has not been established. We assessed agreement between readers of varying experience and a commercial AI algorithm, in DDH detection from infant hip ultrasound sweeps. METHODS: We selected a full spectrum of poor-to-excellent quality images and normal to severe dysplasia, in 240 hips (120 single 2-dimensional images, 120 sweeps). For 12 readers (radiologists, sonographers, clinicians and researchers; 3 were DDH subspecialists), and a ultrasound-FDA-cleared AI software package (Medo Hip), we calculated interobserver reliability for alpha angle measurements by intraclass correlation coefficient (ICC2,1) and for DDH classification by Randolph Kappa. RESULTS: Alpha angle reliability was high for AI versus subspecialists (ICC=0.87 for sweeps, 0.90 for single images). For DDH diagnosis from sweeps, agreement was high between subspecialists (kappa=0.72), and moderate for nonsubspecialists (0.54) and AI (0.47). Agreement was higher for single images (kappa=0.80, 0.66, 0.49). AI reliability deteriorated more than human readers for the poorest-quality images. The agreement of radiologists and clinicians with the accepted standard, while still high, was significantly poorer for sweeps than 2D images (P<0.05). CONCLUSIONS: In a challenging exercise representing the wide spectrum of image quality and reader experience seen in real-world hip ultrasound, agreement on DDH diagnosis from easily obtained sweeps was only slightly lower than from single images, likely because of the additional step of selecting the best image. AI performed similarly to a nonsubspecialist human reader but was more affected by low-quality images.

Diagnostic Accuracy of 3D Ultrasound and Artificial Intelligence for Detection of Pediatric Wrist Injuries.

Wrist trauma is common in children, typically requiring radiography for diagnosis and treatment planning. However, many children do not have fractures and are unnecessarily exposed to radiation. Ultrasound performed at bedside could detect fractures prior to radiography. Modern tools including three-dimensional ultrasound (3DUS) and artificial intelligence (AI) have not yet been applied to this task. Our purpose was to assess (1) feasibility, reliability, and accuracy of 3DUS for detection of pediatric wrist fractures, and (2) accuracy of automated fracture detection via AI from 3DUS sweeps. Children presenting to an emergency department with unilateral upper extremity injury to the wrist region were scanned on both the affected and unaffected limb. Radiographs of the symptomatic limb were obtained for comparison. Ultrasound scans were read by three individuals to determine reliability. An AI network was trained and compared against the human readers. Thirty participants were enrolled, resulting in scans from fifty-five wrists. Readers had a combined sensitivity of 1.00 and specificity of 0.90 for fractures. AI interpretation was indistinguishable from human interpretation, with all fractures detected in the test set of 36 images (sensitivity = 1.0). The high sensitivity of 3D ultrasound and automated AI ultrasound interpretation suggests that ultrasound could potentially rule out fractures in the emergency department.