Deep learning in medical imaging has the potential to minimize the risk of diagnostic errors, reduce radiologist workload, and accelerate diagnosis. Training such deep learning models requires large and accurate datasets, with annotations for all training samples. However, in the medical imaging domain, annotated datasets for specific tasks are often small due to the high complexity of annotations, limited access, or the rarity of diseases. To address this challenge, deep learning models can be pre-trained on large image datasets without annotations using methods from the field of self-supervised learning. After pre-training, small annotated datasets are sufficient to fine-tune the models for a specific task. The most popular self-supervised pre-training approaches in medical imaging are based on contrastive learning. However, recent studies in natural image processing indicate a strong potential for masked autoencoder approaches. Our work compares state-of-the-art contrastive learning methods with the recently introduced masked autoencoder approach “SparK” for convolutional neural networks (CNNs) on medical images. Therefore, we pre-train on a large unannotated CT image dataset and fine-tune on several CT classification tasks. Due to the challenge of obtaining sufficient annotated training data in medical imaging, it is of particular interest to evaluate how the self-supervised pre-training methods perform when fine-tuning on small datasets. By experimenting with gradually reducing the training dataset size for fine-tuning, we find that the reduction has different effects depending on the type of pre-training chosen. The SparK pre-training method is more robust to the training dataset size than the contrastive methods. Based on our results, we propose the SparK pre-training for medical imaging tasks with only small annotated datasets.

Mantle cell lymphoma (MCL) is a rare lymphoid malignancy with a poor prognosis characterised by frequent relapse and short durations of treatment response. Most patients present with aggressive disease, but there exist indolent subtypes without the need for immediate intervention. The very heterogeneous behaviour of MCL is genetically characterised by the translocation t(11;14)(q13;q32), leading to Cyclin D1 overexpression with distinct clinical and biological characteristics and outcomes. There is still an unfulfilled need for precise MCL prognostication in real-time. Machine learning and deep learning neural networks are rapidly advancing technologies with promising results in numerous fields of application. This study develops and compares the performance of deep learning (DL) algorithms and radiomics-based machine learning (ML) models to predict MCL relapse on baseline CT scans. Five classification algorithms were used, including three deep learning models (3D SEResNet50, 3D DenseNet, and an optimised 3D CNN) and two machine learning models based on K-nearest Neighbor (KNN) and Random Forest (RF). The best performing method, our optimised 3D CNN, predicted MCL relapse with a 70% accuracy, better than the 3D SEResNet50 (62%) and the 3D DenseNet (59%). The second-best performing method was the KNN-based machine learning model (64%) after principal component analysis for improved accuracy. Our optimised CNN developed by ourselves correctly predicted MCL relapse in 70% of the patients on baseline CT imaging. Once prospectively tested in clinical trials with a larger sample size, our proposed 3D deep learning model could facilitate clinical management by precision imaging in MCL.

Purpose: Sarcopenia is considered a negative prognostic factor in patients with malignant tumors. Among other diagnostic options, computed tomography (CT), which is repeatedly performed on tumor patients, can be of further benefit. The present study aims to establish a framework for classifying the impact of sarcopenia on the prognosis of patients diagnosed with esophageal or gastric cancer. Additionally, it explores the significance of CT radiomics in both diagnostic and prognostic methodologies. Materials and Methods: CT scans of 83 patients with esophageal or gastric cancer taken at the time of diagnosis and during a follow-up period of one year were evaluated retrospectively. A total of 330 CT scans were analyzed. Seventy three of these patients received operative tumor resection after neoadjuvant chemotherapy, and 74% of the patients were male. The mean age was 64 years (31–83 years). Three time points (t) were defined as a basis for the statistical analysis in order to structure the course of the disease: t1 = initial diagnosis, t2 = following (neoadjuvant) chemotherapy and t3 = end of the first year after surgery in the “surgery” group or end of the first year after chemotherapy. Sarcopenia was determined using the psoas muscle index (PMI). The additional analysis included the analysis of selected radiomic features of the psoas major, quadratus lumborum, and erector spinae muscles at the L3 level. Disease progression was monitored according to the response evaluation criteria in solid tumors (RECIST 1.1). CT scans and radiomics were used to assess the likelihood of tumor progression and their correlation to sarcopenia. For machine learning, the established algorithms decision tree (DT), K-nearest neighbor (KNN), and random forest (RF) were applied. To evaluate the performance of each model, a 10-fold cross-validation as well as a calculation of Accuracy and Area Under the Curve (AUC) was used. Results: During the observation period of the study, there was a significant decrease in PMI. This was most evident in patients with surgical therapy in the comparison between diagnosis and after both neoadjuvant therapy and surgery (each p < 0.001). Tumor progression (PD) was not observed significantly more often in the patients with sarcopenia compared to those without sarcopenia at any time point (p = 0.277 to p = 0.465). On average, PD occurred after 271.69 ± 104.20 days. The time from initial diagnosis to PD in patients “with sarcopenia” was not significantly shorter than in patients “without sarcopenia” at any of the time points (p = 0.521 to p = 0.817). The CT radiomics of skeletal muscle could predict both sarcopenia and tumor progression, with the best results for the psoas major muscle using the RF algorithm. For the detection of sarcopenia, the Accuracy was 0.90 ± 0.03 and AUC was 0.96 ± 0.02. For the prediction of PD, the Accuracy was 0.88 ± 0.04 and the AUC was 0.93 ± 0.04. Conclusions: In the present study, the CT radiomics of skeletal muscle together with machine learning correlated with the presence of sarcopenia, and this can additionally assist in predicting disease progression. These features can be classified as promising alternatives to conventional methods, with great potential for further research and future clinical application. However, when sarcopenia was diagnosed with PMI, no significant correlation between sarcopenia and PD could be observed.

Accurate prediction of lymph node metastasis (LNM) in patients with testicular cancer is highly relevant for treatment decision-making and prognostic evaluation. Our study aimed to develop and validate clinical radiomics models for individual preoperative prediction of LNM in patients with testicular cancer. We enrolled 91 patients with clinicopathologically confirmed early-stage testicular cancer, with disease confined to the testes. We included five significant clinical risk factors (age, preoperative serum tumour markers AFP and B-HCG, histotype and BMI) to build the clinical model. After segmenting 273 retroperitoneal lymph nodes, we then combined the clinical risk factors and lymph node radiomics features to establish combined predictive models using Random Forest (RF), Light Gradient Boosting Machine (LGBM), Support Vector Machine Classifier (SVC), and K-Nearest Neighbours (KNN). Model performance was assessed by the area under the receiver operating characteristic (ROC) curve (AUC). Finally, the decision curve analysis (DCA) was used to evaluate the clinical usefulness. The Random Forest combined clinical lymph node radiomics model with the highest AUC of 0.95 (±0.03 SD; 95% CI) was considered the candidate model with decision curve analysis, demonstrating its usefulness for preoperative prediction in the clinical setting. Our study has identified reliable and predictive machine learning techniques for predicting lymph node metastasis in early-stage testicular cancer. Identifying the most effective machine learning approaches for predictive analysis based on radiomics integrating clinical risk factors can expand the applicability of radiomics in precision oncology and cancer treatment.

A major obstacle to the learning-based segmentation of healthy and tumorous brain tissue is the requirement of having to create a fully labeled training dataset. Obtaining these data requires tedious and error-prone manual labeling with respect to both tumor and non-tumor areas. To mitigate this problem, we propose a new method to obtain high-quality classifiers from a dataset with only small parts of labeled tumor areas. This is achieved by using positive and unlabeled learning in conjunction with a domain adaptation technique. The proposed approach leverages the tumor volume, and we show that it can be either derived with simple measures or completely automatic with a proposed estimation method. While learning from sparse samples allows reducing the necessary annotation time from 4 h to 5 min, we show that the proposed approach further reduces the necessary annotation by roughly 50% while maintaining comparative accuracies compared to traditionally trained classifiers with this approach.

The study’s primary aim is to evaluate the predictive performance of CT-derived 3D radiomics for MCL risk stratification. The secondary objective is to search for radiomic features associated with sustained remission. Included were 70 patients: 31 MCL patients and 39 control subjects with normal axillary lymph nodes followed over five years. Radiomic analysis of all targets (n = 745) was performed and features selected using the Mann Whitney U test; the discriminative power of identifying “high-risk MCL” was evaluated by receiver operating characteristics (ROC). The four radiomic features, “Uniformity”, “Entropy”, “Skewness” and “Difference Entropy” showed predictive significance for relapse (p < 0.05)—in contrast to the routine size measurements, which showed no relevant difference. The best prognostication for relapse achieved the feature “Uniformity” (AUC-ROC-curve 0.87; optimal cut-off ≤0.0159 to predict relapse with 87% sensitivity, 65% specificity, 69% accuracy). Several radiomic features, including the parameter “Short Axis,” were associated with sustained remission. CT-derived 3D radiomics improves the predictive estimation of MCL patients; in combination with the ability to identify potential radiomic features that are characteristic for sustained remission, it may assist physicians in the clinical management of MCL.