Search

Abstract Long COVID is a multi-systemic disease characterized by the persistence or occurrence of many symptoms that in many cases affect the pulmonary system. These, in turn, may deteriorate the patient’s quality of life making it easier to develop severe complications. Being able to predict this syndrome is therefore important as this enables early treatment. In this work, we investigated three machine learning approaches that use clinical data collected at the time of hospitalization to this goal. The first works with all the descriptors feeding a traditional shallow learner, the second exploits the benefits of an ensemble of classifiers, and the third is driven by the intrinsic multimodality of the data so that different models learn complementary information. The experiments on a new cohort of data from 152 patients show that it is possible to predict pulmonary Long Covid sequelae with an accuracy of up to $$94\%$$ 94 % . As a further contribution, this work also publicly discloses the related data repository to foster research in this field.

The “bovine” aortic arch is an anatomic variant consisting in a common origin of the innominate and left carotid artery (CILCA), associated with a greater risk of thoracic aortic diseases (aneurysms and dissections), stroke, and complications after endovascular procedures. CILCA can be detected by visual assessment of computed tomography (CT) chest scans, but it is rarely reported. We developed a deep learning (DL) segmentation-plus-classification system to automatically detect CILCA based on 302 CT studies acquired at 2 centers. One model (3D U-Net) was trained from scratch (supervised by manual segmentation), validated, and tested for the automatic segmentation of the aortic arch and supra-aortic vessels. Three DL architectures (ResNet50, DenseNet-201, and SqueezeNet), pre-trained over millions of common images, were trained, validated, and tested for the automatic classification of CILCA versus non-CILCA, supervised by radiologist’s classification. The 3D U-Net-plus-DenseNet-201 was found to be the best system (Dice index 0.912); its classification performance obtained from internal, independent testing on 126 patients gave a receiver operating characteristic area under the curve of 87.0%, sensitivity 66.7%, specificity 90.5%, positive predictive value 87.5%, negative predictive value 73.1%, positive likelihood ratio 7.0, and negative likelihood ratio 0.4. In conclusion, a combined DL system applied to chest CT scans was developed and proven to be an effective tool to detect individuals with “bovine” aortic arch with a low rate of false-positive findings.

We developed a machine learning model based on radiomics to predict the BI-RADS category of ultrasound-detected suspicious breast lesions and support medical decision-making towards short-interval follow-up versus tissue sampling. From a retrospective 2015–2019 series of ultrasound-guided core needle biopsies performed by four board-certified breast radiologists using six ultrasound systems from three vendors, we collected 821 images of 834 suspicious breast masses from 819 patients, 404 malignant and 430 benign according to histopathology. A balanced image set of biopsy-proven benign (n = 299) and malignant (n = 299) lesions was used for training and cross-validation of ensembles of machine learning algorithms supervised during learning by histopathological diagnosis as a reference standard. Based on a majority vote (over 80% of the votes to have a valid prediction of benign lesion), an ensemble of support vector machines showed an ability to reduce the biopsy rate of benign lesions by 15% to 18%, always keeping a sensitivity over 94%, when externally tested on 236 images from two image sets: (1) 123 lesions (51 malignant and 72 benign) obtained from two ultrasound systems used for training and from a different one, resulting in a positive predictive value (PPV) of 45.9% (95% confidence interval 36.3–55.7%) versus a radiologists’ PPV of 41.5% (p < 0.005), combined with a 98.0% sensitivity (89.6–99.9%); (2) 113 lesions (54 malignant and 59 benign) obtained from two ultrasound systems from vendors different from those used for training, resulting into a 50.5% PPV (40.4–60.6%) versus a radiologists’ PPV of 47.8% (p < 0.005), combined with a 94.4% sensitivity (84.6–98.8%). Errors in BI-RADS 3 category (i.e., assigned by the model as BI-RADS 4) were 0.8% and 2.7% in the Testing set I and II, respectively. The board-certified breast radiologist accepted the BI-RADS classes assigned by the model in 114 masses (92.7%) and modified the BI-RADS classes of 9 breast masses (7.3%). In six of nine cases, the model performed better than the radiologist did, since it assigned a BI-RADS 3 classification to histopathology-confirmed benign masses that were classified as BI-RADS 4 by the radiologist.