Radiomics is an emerging field of image analysis with potential applications in patient risk stratification. This study developed and evaluated machine learning models using quantitative radiomic features extracted from multiparametric magnetic resonance imaging (mpMRI) to detect and classify prostate cancer (PCa). In total, 191 patients that underwent prostatic mpMRI and combined targeted and systematic fusion biopsy were retrospectively included. Segmentations of the whole prostate glands and index lesions were performed manually in apparent diffusion coefficient (ADC) maps and T2-weighted MRI. Radiomic features were extracted from regions corresponding to the whole prostate gland and index lesion. The best performing combination of feature setup and classifier was selected to compare its predictive ability of the radiologist’s evaluation (PI-RADS), mean ADC, prostate specific antigen density (PSAD) and digital rectal examination (DRE) using receiver operating characteristic (ROC) analysis. Models were evaluated using repeated 5-fold cross-validation and a separate independent test cohort. In the test cohort, an ensemble model combining a radiomics model, with models for PI-RADS, PSAD and DRE achieved high predictive AUCs for the differentiation of (i) malignant from benign prostatic lesions (AUC = 0.889) and of (ii) clinically significant (csPCa) from clinically insignificant PCa (cisPCa) (AUC = 0.844). Our combined model was numerically superior to PI-RADS for cancer detection (AUC = 0.779; p = 0.054) as well as for clinical significance prediction (AUC = 0.688; p = 0.209) and showed a significantly better performance compared to mADC for csPCa prediction (AUC = 0.571; p = 0.022). In our study, radiomics accurately characterizes prostatic index lesions and shows performance comparable to radiologists for PCa characterization. Quantitative image data represent a potential biomarker, which, when combined with PI-RADS, PSAD and DRE, predicts csPCa more accurately than mADC. Prognostic machine learning models could assist in csPCa detection and patient selection for MRI-guided biopsy.

In dual-energy CT datasets, the conspicuity of liver metastases can be enhanced by virtual monoenergetic imaging (VMI) reconstructions at low keV levels. Our study investigated whether this effect can be reproduced in photon-counting detector CT (PCD-CT) datasets. We analyzed 100 patients with liver metastases who had undergone contrast-enhanced CT of the abdomen on a PCD-CT (n = 50) or energy-integrating detector CT (EID-CT, single-energy mode, n = 50). PCD-VMI-reconstructions were performed at various keV levels. Identical regions of interest were positioned in metastases, normal liver, and other defined locations assessing image noise, tumor-to-liver ratio (TLR), and contrast-to-noise ratio (CNR). Patients were compared inter-individually. Subgroup analyses were performed according to BMI. On the PCD-CT, noise and CNR peaked at the low end of the keV spectrum. In comparison with the EID-CT, PCD-VMI-reconstructions exhibited lower image noise (at 70 keV) but higher CNR (for ≤70 keV), despite similar CTDIs. Comparing high- and low-BMI patients, CTDI-upregulation was more modest for the PCD-CT but still resulted in similar noise levels and preserved CNR, unlike the EID-CT. In conclusion, PCD-CT VMIs in oncologic patients demonstrated reduced image noise–compared to a standard EID-CT–and improved conspicuity of hypovascularized liver metastases at low keV values. Patients with higher BMIs especially benefited from constant image noise and preservation of lesion conspicuity, despite a more moderate upregulation of CTDI.

Objectives: Structured reporting enhances comparability, readability, and content detail. Large language models (LLMs) could convert free text into structured data without disrupting radiologists’ reporting workflow. This study evaluated an on-premise, privacy-preserving LLM for automatically structuring free-text radiology reports. Materials and methods: We developed an approach to controlling the LLM output, ensuring the validity and completeness of structured reports produced by a locally hosted Llama-2-70B-chat model. A dataset with de-identified narrative chest radiograph (CXR) reports was compiled retrospectively. It included 202 English reports from a publicly available MIMIC-CXR dataset and 197 German reports from our university hospital. Senior radiologist prepared a detailed, fully structured reporting template with 48 question-answer pairs. All reports were independently structured by the LLM and two human readers. Bayesian inference (Markov chain Monte Carlo sampling) was used to estimate the distributions of Matthews correlation coefficient (MCC), with [−0.05, 0.05] as the region of practical equivalence (ROPE). Results: The LLM generated valid structured reports in all cases, achieving an average MCC of 0.75 (94% HDI: 0.70–0.80) and F1 score of 0.70 (0.70–0.80) for English, and 0.66 (0.62–0.70) and 0.68 (0.64–0.72) for German reports, respectively. The MCC differences between LLM and humans were within ROPE for both languages: 0.01 (−0.05 to 0.07), 0.01 (−0.05 to 0.07) for English, and −0.01 (−0.07 to 0.05), 0.00 (−0.06 to 0.06) for German, indicating approximately comparable performance. Conclusion: Locally hosted, open-source LLMs can automatically structure free-text radiology reports with approximately human accuracy. However, the understanding of semantics varied across languages and imaging findings. Key Points: Question Why has structured reporting not been widely adopted in radiology despite clear benefits and how can we improve this? Findings A locally hosted large language model successfully structured narrative reports, showing variation between languages and findings. Critical relevance Structured reporting provides many benefits, but its integration into the clinical routine is limited. Automating the extraction of structured information from radiology reports enables the capture of structured data while allowing the radiologist to maintain their reporting workflow.

The purpose of this study was to evaluate virtual-non contrast reconstructions of Photon-Counting Detector (PCD) CT-angiography datasets using a novel calcium-preserving algorithm (VNC_PC) vs. the standard algorithm (VNC_Conv) for their potential to replace unenhanced acquisitions (TNC) in patients after endovascular aneurysm repair (EVAR). 20 EVAR patients who had undergone CTA (unenhanced and arterial phase) on a novel PCD-CT were included. VNC_Conv- and VNC_PC-series were derived from CTA-datasets and intraluminal signal and noise compared. Three readers evaluated image quality, contrast removal, and removal of calcifications/stent parts and assessed all VNC-series for their suitability to replace TNC-series. Image noise was higher in VNC- than in TNC-series (18.6 ± 5.3 HU, 16.7 ± 7.1 HU, and 14.9 ± 7.1 HU for VNC_Conv-, VNC_PC-, and TNC-series, p = 0.006). Subjective image quality was substantially higher in VNC_PC- than VNC_Conv-series (4.2 ± 0.9 vs. 2.5 ± 0.6; p < 0.001). Aortic contrast removal was complete in all VNC-series. Unlike in VNC_Conv-reconstructions, only minuscule parts of stents or calcifications were erroneously subtracted in VNC_PC-reconstructions. Readers considered 95% of VNC_PC-series fully or mostly suited to replace TNC-series; for VNC_Conv-reconstructions, however, only 75% were considered mostly (and none fully) suited for TNC-replacement. VNC_PC-reconstructions of PCD-CT-angiography datasets have excellent image quality with complete contrast removal and only minimal erroneous subtractions of stent parts/calcifications. They could replace TNC-series in almost all cases.

Objectives: The goal of this study is to demonstrate the performance of radiomics and CNN-based classifiers in determining the primary origin of gastrointestinal liver metastases for visually indistinguishable lesions. Methods: In this retrospective, IRB-approved study, 31 pancreatic cancer patients with 861 lesions (median age [IQR]: 65.39 [56.87, 75.08], 48.4% male) and 47 colorectal cancer patients with 435 lesions (median age [IQR]: 65.79 [56.99, 74.62], 63.8% male) were enrolled. A pretrained nnU-Net performed automated segmentation of 1296 liver lesions. Radiomics features for each lesion were extracted using pyradiomics. The performance of several radiomics-based machine-learning classifiers was investigated for the lesions and compared to an image-based deep-learning approach using a DenseNet-121. The performance was evaluated by AUC/ROC analysis. Results: The radiomics-based K-nearest neighbor classifier showed the best performance on an independent test set with AUC values of 0.87 and an accuracy of 0.67. In comparison, the image-based DenseNet-121-classifier reached an AUC of 0.80 and an accuracy of 0.83. Conclusions: CT-based radiomics and deep learning can distinguish the etiology of liver metastases from gastrointestinal primary tumors. Compared to deep learning, radiomics based models showed a varying generalizability in distinguishing liver metastases from colorectal cancer and pancreatic adenocarcinoma.