Gallium-68 prostate-specific membrane antigen positron emission tomography (⁶⁸Ga-PSMA-PET) is a highly sensitive method to detect prostate cancer (PC) metastases. Visual discrimination between malignant and physiologic/unspecific tracer accumulation by a nuclear medicine (NM) specialist is essential for image interpretation. In the future, automated machine learning (ML)-based tools will assist physicians in image analysis. The aim of this work was to develop a tool for analysis of ⁶⁸Ga-PSMA-PET images and to compare its efficacy to that of human readers. Five different ML methods were compared and tested on multiple positron emission tomography/computed tomography (PET/CT) data-sets. Forty textural features extracted from both PET- and low-dose CT data were analyzed. In total, 2419 hotspots from 72 patients were included. Comparing results from human readers to those of ML-based analyses, up to 98% area under the curve (AUC), 94% sensitivity (SE), and 89% specificity (SP) were achieved. Interestingly, textural features assessed in native low-dose CT increased the accuracy significantly. Thus, ML based on ⁶⁸Ga-PSMA-PET/CT radiomics features can classify hotspots with high precision, comparable to that of experienced NM physicians. Additionally, the superiority of multimodal ML-based analysis considering all PET and low-dose CT features was shown. Morphological features seemed to be of special additional importance even though they were extracted from native low-dose CTs.

The importance of machine learning (ML) in the clinical environment increases constantly. Differentiation of pathological from physiological tracer-uptake in positron emission tomography/computed tomography (PET/CT) images is considered time-consuming and attention intensive, hence crucial for diagnosis and treatment planning. This study aimed at comparing and validating supervised ML algorithms to classify pathological uptake in prostate cancer (PC) patients based on prostate-specific membrane antigen (PSMA)-PET/CT. Retrospective analysis of ⁶⁸Ga-PSMA-PET/CTs of 72 PC patients resulted in a total of 77 radiomics features from 2452 manually delineated hotspots for training and labeled pathological (1629) or physiological (823) as ground truth (GT). As the held-out test dataset, 331 hotspots (path.:128, phys.: 203) were delineated in 15 other patients. Three ML classifiers were trained and ranked to assess classification performance. As a result, a high overall average performance (area under the curve (AUC) of 0.98) was achieved, especially to detect pathological uptake (0.97 mean sensitivity). However, there is still room for improvement to detect physiological uptake (0.82 mean specificity), especially for glands. The ML algorithm applied to manually delineated lesions predicts hotspot labels with high accuracy on unseen data and may be an important tool to assist in clinical diagnosis.

Peripheral arterial disease (PAD) is caused by atherosclerosis and is a common disease of the elderly leading to excess morbidity and mortality. Early PAD diagnosis is important, as the only available causal therapy is addressing risk factors like smoking, hypercholesterolemia or hypertension. However, current diagnostic techniques often do not detect early stages of PAD. We theorize that PAD’s underlying cause atherosclerosis can be detected on color fundus photography (CFP) images with a convolutional neural network architecture, which might aid earlier PAD diagnosis and improve disease monitoring. In this explorative study a deep attention-based Multiple Instance Learning (MIL) architecture is used to capture retinal imaging biomarkers on CFP images of 135 examinations. To capture subtle variations in vascular structures, higher image resolution can be utilized by partitioning the CFP into patches. Our architecture converts each patch into a feature vector, and determines its relative importance via an automatically computed attention weight. Our best model achieves an ROC AUC score of 0.890. Visualizing these attention weights provides insights about the network’s decision and suggests ocular involvement in PAD. Statistical analysis confirms that the optic disc and the temporal arcades are weighted significantly higher (p < 0.001) than retinal background. Our results support the feasibility of detecting the presence of PAD with a modern deep learning approach.