Purpose: To classify COVID-19, COVID-19-like and non-COVID-19 interstitial pneumonia using lung CT radiomic features. Material and Methods: CT data of 115 patients with respiratory symptoms suspected for COVID-19 disease were retrospectively analyzed. Based on the results of nasopharyngeal swab, patients were divided into two main groups, COVID-19 positive (C +) and COVID-19 negative (C−), respectively. C− patients, however, presented with interstitial lung involvement. A subgroup of C−, COVID-19-like (CL), were considered as highly suggestive of COVID pneumonia at CT. Radiomic features were extracted from the whole lungs. A dual machine learning (ML) model approach was used. The first one excluded CL patients from the training set, eventually included on the test set. The second model included the CL patients also in the training set. Results: The first model classified C + and C− pneumonias with AUC of 0.83. CL median response (0.80) was more similar to C + (0.92) compared to C− (0.17). Radiomic footprints of CL were similar to the C + ones (possibly false negative swab test). The second model, however, merging C + with CL patients in the training set, showed a slight decrease in classification performance (AUC = 0.81). Conclusion: Whole lung ML models based on radiomics can classify C + and C− interstitial pneumonia. This may help in the correct management of patients with clinical and radiological stigmata of COVID-19, however presenting with a negative swab test. CL pneumonia was similar to C + pneumonia, albeit with slightly different radiomic footprints.

Objectives: This study evaluated the feasibility of DWI for lesion targeting in MRI-guided breast biopsies. Furthermore, it assessed device positioning on DWI during biopsy procedures. Methods: A total of 87 biopsy procedures (5/87 bilateral) consecutively performed between March 2019 and June 2020 were retrospectively reviewed: in these procedures, a preliminary DWI sequence (b = 1300 s/mm²) was acquired to assess lesion detectability. We included 64/87 procedures on lesions detectable at DWI; DWI sequences were added to the standard protocol to localize lesion and biopsy device and to assess the site marker correct positioning. Results: Mass lesions ranged from 5 to 48 mm, with a mean size of 10.7 mm and a median size of 8 mm. Non-mass lesions ranged from 7 to 90 mm, with a mean size of 33.9 mm and a median size of 31 mm. Positioning of the coaxial system was confirmed on both T1-weighted and DWI sequences. At DWI, the biopsy needle was detectable in 62/64 (96.9%) cases; it was not visible in 2/64 (3.1%) cases. The site marker was always identified using T1-weighted imaging; a final DWI sequence was acquired in 44/64 cases (68.8%). In 42/44 cases (95.5%), the marker was recognizable at DWI. Conclusions: DWI can be used as a cost-effective, highly reliable technique for targeting both mass and non-mass lesions, with a minimum size of 5 mm, detectable at pre-procedural DWI. DWI is also a feasible technique to localize the biopsy device and to confirm the deployment of the site marker. Key Points: • MRI-guided breast biopsy is performed in referral centers by an expert dedicated staff, based on prior MR imaging; contrast agent administration is usually needed for lesion targeting. • DWI represents a feasible, highly reliable technique for lesion targeting, avoiding contrast agent administration. • DWI allows a precise localization of both biopsy needle device and site marker.

In the era of precision medicine, increasing importance is given to machine learning (ML) applications. In breast cancer, advanced analyses, such as the radiomic process, characterise tumours and predict therapy responses. Breast magnetic resonance imaging (MRI) plays a key role in screening, staging, and treatment monitoring. Lesion segmentation on MRI is essential both to assess tumour growth and as a baseline for radiomic feature extraction. Manual segmentation is time-consuming and prone to inter-operator variability, limiting access to large labelled datasets and robust analyses. The use of ML for breast lesion segmentation on MRI has been investigated through a systematic review of PubMed, exploring studies published over the last 10 years. Results are compared in terms of performance, primarily using the Dice score. Early unsupervised methods achieved a mean Dice score of ∼0.75, surpassing traditional supervised methods (∼0.70). In contrast, deep learning (DL) approaches based on U-Net achieved higher average scores of 0.79. Further customised supervised DL approaches reached a mean Dice score of ∼0.83. However, there is still a gap in research on unsupervised DL techniques, which could help reduce bias and human variability. Future work may also explore multiparametric and multitechnique data, integrating more representative samples, including non-mass lesions.