Juan Miguel Valverde

Juan Miguel Valverde, Artem Shatillo, Jussi Tohka

We introduce Sauron, a filter pruning method that eliminates redundant feature maps of convolutional neural networks (CNNs). Sauron optimizes, jointly with the loss function, a regularization term that promotes feature maps clustering at each convolutional layer by reducing the distance between feature maps. Sauron then eliminates the filters corresponding to the redundant feature maps by using automatically adjusted layer-specific thresholds. Unlike most filter pruning methods, Sauron requires minimal changes to typical neural network optimization because it prunes and optimizes CNNs jointly, which, in turn, accelerates the optimization over time. Moreover, unlike with other cluster-based approaches, the user does not need to specify the number of clusters in advance, a hyperparameter that is difficult to tune. We evaluated Sauron and five state-of-the-art filter pruning methods on four medical image segmentation tasks. This is an area where little attention has been paid to filter pruning, but where smaller CNN models are desirable for local deployment, mitigating privacy concerns associated with cloud-based solutions. Sauron was the only method that achieved a reduction in model size of over 90% without deteriorating substantially the performance. Sauron also achieved, overall, the fastest models at inference time in machines with and without GPUs. Finally, we show through experiments that the feature maps of models pruned with Sauron are highly interpretable, which is essential for medical image segmentation.

Kevin Wilkinghoff, Neelu Madan, Juan Miguel Valverde et al.

Anomaly detection aims to identify observations that deviate from expected behavior. Because anomalous events are inherently sparse, most frameworks are trained exclusively on normal data to learn a single reference model of normality. This implicitly assumes that normal behavior can be captured by a single, unconditional reference distribution. In practice, however, anomalies are often context-dependent: A specific observation may be normal under one operating condition, yet anomalous under another. As machine learning systems are deployed in dynamic and heterogeneous environments, these fixed-context assumptions introduce structural ambiguity, i.e., the inability to distinguish contextual variation from genuine abnormality under marginal modeling, leading to unstable performance and unreliable anomaly assessments. While modern sensing systems frequently collect multimodal data capturing complementary aspects of both system behavior and operating conditions, existing methods treat all data streams equally, without distinguishing contextual information from anomaly-relevant signals. As a result, abnormality is often evaluated without explicitly conditioning on operating conditions. We argue that multimodal anomaly detection should be reframed as a cross-modal contextual inference problem, in which modalities play asymmetric roles, separating context from observation, to define abnormality conditionally rather than relative to a single global reference. This perspective has implications for model design, evaluation protocols, and benchmark construction, and outline open research challenges toward robust, context-aware multimodal anomaly detection.

Juan Miguel Valverde, Dim P. Papadopoulos, Rasmus Larsen et al.

Standard deep learning models for image segmentation cannot guarantee topology accuracy, failing to preserve the correct number of connected components or structures. This, in turn, affects the quality of the segmentations and compromises the reliability of the subsequent quantification analyses. Previous works have proposed to enhance topology accuracy with specialized frameworks, architectures, and loss functions. However, these methods are often cumbersome to integrate into existing training pipelines, they are computationally very expensive, or they are restricted to structures with tubular morphology. We present SCNP, an efficient method that improves topology accuracy by penalizing the logits with their poorest-classified neighbor, forcing the model to improve the prediction at the pixels' neighbors before allowing it to improve the pixels themselves. We show the effectiveness of SCNP across 13 datasets, covering different structure morphologies and image modalities, and integrate it into three frameworks for semantic and instance segmentation. Additionally, we show that SCNP can be integrated into several loss functions, making them improve topology accuracy. Our code can be found at https://jmlipman.github.io/SCNP-SameClassNeighborPenalization.

Juan Miguel Valverde, Motoya Koga, Nijihiko Otsuka et al.

We present TopoMortar, a brick wall dataset that is the first dataset specifically designed to evaluate topology-focused image segmentation methods, such as topology loss functions. Motivated by the known sensitivity of methods to dataset challenges, such as small training sets, noisy labels, and out-of-distribution test-set images, TopoMortar is created to enable in two ways investigating methods' effectiveness at improving topology accuracy. First, by eliminating dataset challenges that, as we show, impact the effectiveness of topology loss functions. Second, by allowing to represent different dataset challenges in the same dataset, isolating methods' performance from dataset challenges. TopoMortar includes three types of labels (accurate, pseudo-labels, and noisy labels), two fixed training sets (large and small), and in-distribution and out-of-distribution test-set images. We compared eight loss functions on TopoMortar, and we found that clDice achieved the most topologically accurate segmentations, and that the relative advantageousness of the other loss functions depends on the experimental setting. Additionally, we show that data augmentation and self-distillation can elevate Cross entropy Dice loss to surpass most topology loss functions, and that those simple methods can enhance topology loss functions as well. TopoMortar and our code can be found at https://jmlipman.github.io/TopoMortar

Juan Miguel Valverde, Maja Østergaard, Adrian Rodriguez-Palomo et al.

Accurate segmentation of thin, tubular structures (e.g., blood vessels) is challenging for deep neural networks. These networks classify individual pixels, and even minor misclassifications can break the thin connections within these structures. Existing methods for improving topology accuracy, such as topology loss functions, rely on very precise, topologically-accurate training labels, which are difficult to obtain. This is because annotating images, especially 3D images, is extremely laborious and time-consuming. Low image resolution and contrast further complicates the annotation by causing tubular structures to appear disconnected. We present CoLeTra, a data augmentation strategy that integrates to the models the prior knowledge that structures that appear broken are actually connected. This is achieved by creating images with the appearance of disconnected structures while maintaining the original labels. Our extensive experiments, involving different architectures, loss functions, and datasets, demonstrate that CoLeTra leads to segmentations topologically more accurate while often improving the Dice coefficient and Hausdorff distance. CoLeTra's hyper-parameters are intuitive to tune, and our sensitivity analysis shows that CoLeTra is robust to changes in these hyper-parameters. We also release a dataset specifically suited for image segmentation methods with a focus on topology accuracy. CoLetra's code can be found at https://github.com/jmlipman/CoLeTra.

Juan Miguel Valverde, Artem Shatillo, Riccardo de Feo et al.

We present MedicDeepLabv3+, a convolutional neural network that is the first completely automatic method to segment cerebral hemispheres in magnetic resonance (MR) volumes of rats with lesions. MedicDeepLabv3+ improves the state-of-the-art DeepLabv3+ with an advanced decoder, incorporating spatial attention layers and additional skip connections that, as we show in our experiments, lead to more precise segmentations. MedicDeepLabv3+ requires no MR image preprocessing, such as bias-field correction or registration to a template, produces segmentations in less than a second, and its GPU memory requirements can be adjusted based on the available resources. We optimized MedicDeepLabv3+ and six other state-of-the-art convolutional neural networks (DeepLabv3+, UNet, HighRes3DNet, V-Net, VoxResNet, Demon) on a heterogeneous training set comprised by MR volumes from 11 cohorts acquired at different lesion stages. Then, we evaluated the trained models and two approaches specifically designed for rodent MRI skull stripping (RATS and RBET) on a large dataset of 655 MR rat brain volumes. In our experiments, MedicDeepLabv3+ outperformed the other methods, yielding an average Dice coefficient of 0.952 and 0.944 in the brain and contralateral hemisphere regions. Additionally, we show that despite limiting the GPU memory and the training data, our MedicDeepLabv3+ also provided satisfactory segmentations. In conclusion, our method, publicly available at https://github.com/jmlipman/MedicDeepLabv3Plus, yielded excellent results in multiple scenarios, demonstrating its capability to reduce human workload in rat neuroimaging studies.

Juan Miguel Valverde, Jussi Tohka

We propose Region-wise (RW) loss for biomedical image segmentation. Region-wise loss is versatile, can simultaneously account for class imbalance and pixel importance, and it can be easily implemented as the pixel-wise multiplication between the softmax output and a RW map. We show that, under the proposed RW loss framework, certain loss functions, such as Active Contour and Boundary loss, can be reformulated similarly with appropriate RW maps, thus revealing their underlying similarities and a new perspective to understand these loss functions. We investigate the observed optimization instability caused by certain RW maps, such as Boundary loss distance maps, and we introduce a mathematically-grounded principle to avoid such instability. This principle provides excellent adaptability to any dataset and practically ensures convergence without extra regularization terms or optimization tricks. Following this principle, we propose a simple version of boundary distance maps called rectified Region-wise (RRW) maps that, as we demonstrate in our experiments, achieve state-of-the-art performance with similar or better Dice coefficients and Hausdorff distances than Dice, Focal, weighted Cross entropy, and Boundary losses in three distinct segmentation tasks. We quantify the optimization instability provided by Boundary loss distance maps, and we empirically show that our RRW maps are stable to optimize. The code to run all our experiments is publicly available at: https://github.com/jmlipman/RegionWiseLoss.

Juan Miguel Valverde, Artem Shatillo, Riccardo de Feo et al.

We present a fully convolutional neural network (ConvNet), named RatLesNetv2, for segmenting lesions in rodent magnetic resonance (MR) brain images. RatLesNetv2 architecture resembles an autoencoder and it incorporates residual blocks that facilitate its optimization. RatLesNetv2 is trained end to end on three-dimensional images and it requires no preprocessing. We evaluated RatLesNetv2 on an exceptionally large dataset composed of 916 T2-weighted rat brain MRI scans of 671 rats at nine different lesion stages that were used to study focal cerebral ischemia for drug development. In addition, we compared its performance with three other ConvNets specifically designed for medical image segmentation. RatLesNetv2 obtained similar to higher Dice coefficient values than the other ConvNets and it produced much more realistic and compact segmentations with notably fewer holes and lower Hausdorff distance. The Dice scores of RatLesNetv2 segmentations also exceeded inter-rater agreement of manual segmentations. In conclusion, RatLesNetv2 could be used for automated lesion segmentation, reducing human workload and improving reproducibility. RatLesNetv2 is publicly available at https://github.com/jmlipman/RatLesNetv2.

Juan Miguel Valverde, Vandad Imani, Ali Abdollahzadeh et al.

Transfer learning refers to machine learning techniques that focus on acquiring knowledge from related tasks to improve generalization in the tasks of interest. In MRI, transfer learning is important for developing strategies that address the variation in MR images. Additionally, transfer learning is beneficial to re-utilize machine learning models that were trained to solve related tasks to the task of interest. Our goal is to identify research directions, gaps of knowledge, applications, and widely used strategies among the transfer learning approaches applied in MR brain imaging. We performed a systematic literature search for articles that applied transfer learning to MR brain imaging. We screened 433 studies and we categorized and extracted relevant information, including task type, application, and machine learning methods. Furthermore, we closely examined brain MRI-specific transfer learning approaches and other methods that tackled privacy, unseen target domains, and unlabeled data. We found 129 articles that applied transfer learning to brain MRI tasks. The most frequent applications were dementia related classification tasks and brain tumor segmentation. A majority of articles utilized transfer learning on convolutional neural networks (CNNs). Only few approaches were clearly brain MRI specific, considered privacy issues, unseen target domains or unlabeled data. We proposed a new categorization to group specific, widely-used approaches. There is an increasing interest in transfer learning within brain MRI. Public datasets have contributed to the popularity of Alzheimer's diagnostics/prognostics and tumor segmentation. Likewise, the availability of pretrained CNNs has promoted their utilization. Finally, the majority of the surveyed studies did not examine in detail the interpretation of their strategies after applying transfer learning, and did not compare to other approaches.

Juan Miguel Valverde, Vandad Imani, John D. Lewis et al.

We propose a four-layer fully-connected neural network (FNN) for predicting fluid intelligence scores from T1-weighted MR images for the ABCD-challenge. In addition to the volumes of brain structures, the FNN uses cortical WM/GM contrast and cortical thickness at 78 cortical regions. These last two measurements were derived from the T1-weighted MR images using cortical surfaces produced by the CIVET pipeline. The age and gender of the subjects and the scanner manufacturer are also used as features for the learning algorithm. This yielded 283 features provided to the FNN with two hidden layers of 20 and 15 nodes. The method was applied to the data from the ABCD study. Trained with a training set of 3736 subjects, the proposed method achieved a MSE of 71.596 and a correlation of 0.151 in the validation set of 415 subjects. For the final submission, the model was trained with 3568 subjects and it achieved a MSE of 94.0270 in the test set comprised of 4383 subjects.

Juan Miguel Valverde, Artem Shatillo, Riccardo de Feo et al.

Manual segmentation of rodent brain lesions from magnetic resonance images (MRIs) is an arduous, time-consuming and subjective task that is highly important in pre-clinical research. Several automatic methods have been developed for different human brain MRI segmentation, but little research has targeted automatic rodent lesion segmentation. The existing tools for performing automatic lesion segmentation in rodents are constrained by strict assumptions about the data. Deep learning has been successfully used for medical image segmentation. However, there has not been any deep learning approach specifically designed for tackling rodent brain lesion segmentation. In this work, we propose a novel Fully Convolutional Network (FCN), RatLesNet, for the aforementioned task. Our dataset consists of 131 T2-weighted rat brain scans from 4 different studies in which ischemic stroke was induced by transient middle cerebral artery occlusion. We compare our method with two other 3D FCNs originally developed for anatomical segmentation (VoxResNet and 3D-U-Net) with 5-fold cross-validation on a single study and a generalization test, where the training was done on a single study and testing on three remaining studies. The labels generated by our method were quantitatively and qualitatively better than the predictions of the compared methods. The average Dice coefficient achieved in the 5-fold cross-validation experiment with the proposed approach was 0.88, between 3.7% and 38% higher than the compared architectures. The presented architecture also outperformed the other FCNs at generalizing on different studies, achieving the average Dice coefficient of 0.79.