CMIV Publications

Many inherently ambiguous tasks in medical imaging suffer from inter-observer variability, resulting in a reference standard defined by a distribution of labels with high variance. Training only on a consensus or majority vote label, as is common in medical imaging, discards valuable information on uncertainty amongst a panel of experts. In this work, we propose to train on the full label distribution to predict the uncertainty within a panel of experts and the most likely ground-truth label. To do so, we propose a new stochastic classification framework based on the conditional variational auto-encoder, which we refer to as the Latent Doctor Model (LDM). In an extensive comparative analysis, we compare the LDM with a model trained on the majority vote label and other methods capable of learning a distribution of labels. We show that the LDM is able to reproduce the reference-standard distribution significantly better than the majority vote baseline. Compared to the other baseline methods, we demonstrate that the LDM performs best at modeling the label distribution and its corresponding uncertainty in two prostate tumor grading tasks. Furthermore, we show competitive performance of the LDM with the more computationally demanding deep ensembles on a tumor budding classification task.

Background: Quality of life (QoL) is affected up to 5 years after an anterior cruciate ligament (ACL) injury. Knee impairment and osteoarthritis (OA) development increase over time, and this may affect QoL at a long-term follow-up.Purpose: To investigate changes in health- and knee-related QoL between 20 and 35 years after ACL injury and compare it between patients treated with or without ACL surgery, as well as to study how symptomatic OA (SOA) is associated with change in QoL.Study Design: Cohort study; Level of evidence, 2.Methods: Between 1980 and 1983, 139 patients with acute ACL rupture were allocated to surgical or nonsurgical treatment of the ACL. Both groups completed a structured rehabilitation program. Of those patients, 59 were followed for 20 and 35 years after ACL rupture. After 10 crossovers, 33 patients were treated with primary repair or ACL reconstruction, and 26 were treated without ACL surgery. Combined radiographic OA and knee symptoms at 35 years was defined as SOA. QoL was assessed at 20 and 35 years after injury with the Knee injury and Osteoarthritis Outcome Score QoL (KOOS-QoL) subscale (range, 1-100), ACL-QoL questionnaire (total score and 5 subscales; range, 1-100), European QoL-5 Dimensions Questionnaire, and visual analog scale. Results were analyzed with paired and independent-sample t tests and chi-square tests.Results: Knee-related QoL was impaired at both 20 and 35 years after ACL injury, and differences were dependent on the measurement outcome. In the total cohort, KOOS-QoL did not change but both total ACL-QoL score (7.1 points; 95% CI, 2.2-11.9) and 4 of 5 subscales (5-10 points) decreased (P < .05). No differences were found between treatment groups. QoL decreased overall in patients with SOA, with a 21-point difference within-group change in KOOS-QoL (SOA or non-SOA) between 20 and 35 years of follow-up (P = .001; Cohen d = 1.0).Conclusion: An ACL injury impairs knee-related QoL for up to 35 years, with no difference between treatment approaches (initial repair or later reconstruction compared with nonsurgical treatment). The deterioration decreases with longer follow-up. Clinicians should be aware of differences in QoL depending on the measurement outcome.

The ability to detect anomalies, i.e. anything not seen during training or out -of -distribution (OOD), in medical imaging applications is essential for successfully deploying machine learning systems. Filtering out OOD data using unsupervised learning is especially promising because it does not require costly annotations. A new class of models called AnoDDPMs, based on denoising diffusion probabilistic models (DDPMs), has recently achieved significant progress in unsupervised OOD detection. This work provides a benchmark for unsupervised OOD detection methods in digital pathology. By leveraging fast sampling techniques, we apply AnoDDPM on a large enough scale for whole -slide image analysis on the complete test set of the Camelyon16 challenge. Based on ROC analysis, we show that AnoDDPMs can detect OOD data with an AUC of up to 94.13 and 86.93 on two patch -level OOD detection tasks, outperforming the other unsupervised methods. We observe that AnoDDPMs alter the semantic properties of inputs, replacing anomalous data with more benign -looking tissue. Furthermore, we highlight the flexibility of AnoDDPM towards different information bottlenecks by evaluating reconstruction errors for inputs with different signal-to-noise ratios. While there is still a significant performance gap with fully supervised learning, AnoDDPMs show considerable promise in the field of OOD detection in digital pathology.