Loss Weightings for Improving Imbalanced Brain Structure Segmentation Using Fully Convolutional Networks

Volumetric Breast Density Estimation From Three-Dimensional Reconstructed Digital Breast Tomosynthesis Images Using Deep Learning

PURPOSE

Breast density is a widely established independent breast cancer risk factor. With the increasing utilization of digital breast tomosynthesis (DBT) in breast cancer screening, there is an opportunity to estimate volumetric breast density (VBD) routinely. However, current available methods extrapolate VBD from two-dimensional (2D) images acquired using DBT and/or depend on the existence of raw DBT data, which is rarely archived by clinical centers because of storage constraints.

METHODS

We retrospectively analyzed 1,080 nonactionable three-dimensional (3D) reconstructed DBT screening examinations acquired between 2011 and 2016. Reference tissue segmentations were generated using previously validated software that uses 3D reconstructed slices and raw 2D DBT data. We developed a deep learning (DL) model that segments dense and fatty breast tissue from background. We then applied this model to estimate %VBD and absolute dense volume (ADV) in cm3 in a separate case-control sample (180 cases and 654 controls). We created two conditional logistic regression models, relating each model-derived density measurement to likelihood of contralateral breast cancer diagnosis, adjusted for age, BMI, family history, and menopausal status.

RESULTS

The DL model achieved unweighted and weighted Dice scores of 0.88 (standard deviation [SD] = 0.08) and 0.76 (SD = 0.15), respectively, on the held-out test set, demonstrating good agreement between the model and 3D reference segmentations. There was a significant association between the odds of breast cancer diagnosis and model-derived VBD (odds ratio [OR], 1.41 [95 % CI, 1.13 to 1.77]; P = .002), with an AUC of 0.65 (95% CI, 0.60 to 0.69). ADV was also significantly associated with breast cancer diagnosis (OR, 1.45 [95% CI, 1.22 to 1.73]; P < .001) with an AUC of 0.67 (95% CI, 0.62 to 0.71).

CONCLUSION

DL-derived density measures derived from 3D reconstructed DBT images are associated with breast cancer diagnosis.

Efficient In-Training Adaptive Compound Loss Function Contribution Control for Medical Image Segmentation

Image segmentation plays a crucial role in many clinical applications, including disease diagnosis and monitoring. Current state-of-the-art segmentation approaches use deep neural networks that are trained on their target tasks by minimizing a loss function. Class imbalance is one of the major challenges that these networks face, where the target object is significantly underrepresented. Compound loss functions that incorporate the binary cross-entropy (BCE) and Dice loss are among the most prominent approaches to address this issue. However, determining the contribution of each individual loss to the overall compound loss function is a tedious process. It requires hyperparameter fine-tuning and multiple iterations of training, which is highly inefficient in terms of time and energy consumption. To address this issue, we propose an approach that adaptively controls the contribution of each of these individual loss functions during training. This eliminates the need for multiple fine-tuning iterations to achieve the desired precision and recall for segmentation models.

Projected pooling loss for red nucleus segmentation with soft topology constraints

Purpose

Deep learning is the standard for medical image segmentation. However, it may encounter difficulties when the training set is small. Also, it may generate anatomically aberrant segmentations. Anatomical knowledge can be potentially useful as a constraint in deep learning segmentation methods. We propose a loss function based on projected pooling to introduce soft topological constraints. Our main application is the segmentation of the red nucleus from quantitative susceptibility mapping (QSM) which is of interest in parkinsonian syndromes.

Approach

This new loss function introduces soft constraints on the topology by magnifying small parts of the structure to segment to avoid that they are discarded in the segmentation process. To that purpose, we use projection of the structure onto the three planes and then use a series of MaxPooling operations with increasing kernel sizes. These operations are performed both for the ground truth and the prediction and the difference is computed to obtain the loss function. As a result, it can reduce topological errors as well as defects in the structure boundary. The approach is easy to implement and computationally efficient.

Results

When applied to the segmentation of the red nucleus from QSM data, the approach led to a very high accuracy (Dice 89.9%) and no topological errors. Moreover, the proposed loss function improved the Dice accuracy over the baseline when the training set was small. We also studied three tasks from the medical segmentation decathlon challenge (MSD) (heart, spleen, and hippocampus). For the MSD tasks, the Dice accuracies were similar for both approaches but the topological errors were reduced.

Conclusions

We propose an effective method to automatically segment the red nucleus which is based on a new loss for introducing topology constraints in deep learning segmentation.

Improving the Generalizability of Deep Learning for T2-Lesion Segmentation of Gliomas in the Post-Treatment Setting

Although fully automated volumetric approaches for monitoring brain tumor response have many advantages, most available deep learning models are optimized for highly curated, multi-contrast MRI from newly diagnosed gliomas, which are not representative of post-treatment cases in the clinic. Improving segmentation for treated patients is critical to accurately tracking changes in response to therapy. We investigated mixing data from newly diagnosed (n = 208) and treated (n = 221) gliomas in training, applying transfer learning (TL) from pre- to post-treatment imaging domains, and incorporating spatial regularization for T2-lesion segmentation using only T2 FLAIR images as input to improve generalization post-treatment. These approaches were evaluated on 24 patients suspected of progression who had received prior treatment. Including 26% of treated patients in training improved performance by 13.9%, and including more treated and untreated patients resulted in minimal changes. Fine-tuning with treated glioma improved sensitivity compared to data mixing by 2.5% (p < 0.05), and spatial regularization further improved performance when used with TL by 95th HD, Dice, and sensitivity (6.8%, 0.8%, 2.2%; p < 0.05). While training with ≥60 treated patients yielded the majority of performance gain, TL and spatial regularization further improved T2-lesion segmentation to treated gliomas using a single MR contrast and minimal processing, demonstrating clinical utility in response assessment.

Improved segmentation of basal ganglia from MR images using convolutional neural network with crossover-typed skip connection

PURPOSE

Accurate and automatic segmentation of basal ganglia from magnetic resonance (MR) images is important for diagnosis and treatment of various brain disorders. However, the basal ganglia segmentation is a challenging task because of the class imbalance and the unclear boundaries among basal ganglia anatomical structures. Thus, we aim to present an encoder-decoder convolutional neural network (CNN)-based method for improved segmentation of basal ganglia by focusing on skip connections that determine the segmentation performance of encoder-decoder CNNs. We also aim to reveal the effect of skip connections on the segmentation of basal ganglia with unclear boundaries.

METHODS

We used the encoder-decoder CNNs with the following five patterns of skip connections: without skip connection, with full-resolution horizontal skip connection, with horizontal skip connections, with vertical skip connections, and with crossover-typed skip connections (the proposed method). We compared and evaluated the performance of the CNNs in the experiment of basal ganglia segmentation using T1-weighted MR brain images of 79 patients.

RESULTS

The experimental results showed that the skip connections at each scale level help CNNs to acquire multi-scale image features, the vertical skip connections contribute on acquiring finer image features for segmentation of smaller anatomical structures with more blurred boundaries, and the crossover-typed skip connections, a combination of horizontal and vertical skip connections, provided better segmentation accuracy.

CONCLUSION

This paper investigated the effect of skip connections on the basal ganglia segmentation and revealed the crossover-typed skip connections might be effective for improving the segmentation of basal ganglia with the class imbalance and the unclear boundaries.

Interpretable Skin Cancer Classification based on Incremental Domain Knowledge Learning

The recent advances in artificial intelligence have led to the rapid development of computer-aided skin cancer diagnosis applications that perform on par with dermatologists. However, the black-box nature of such applications makes it difficult for physicians to trust the predicted decisions, subsequently preventing the proliferation of such applications in the clinical workflow. In this work, we aim to address this challenge by developing an interpretable skin cancer diagnosis approach using clinical images. Accordingly, a skin cancer diagnosis model consolidated with two interpretability methods is developed. The first interpretability method integrates skin cancer diagnosis domain knowledge, characterized by a skin lesion taxonomy, into model development, whereas the other method focuses on visualizing the decision-making process by highlighting the dominant of interest regions of skin lesion images. The proposed model is trained and validated on clinical images since the latter are easily obtainable by non-specialist healthcare providers. The results demonstrate the effectiveness of incorporating lesion taxonomy in improving model classification accuracy, where our model can predict the skin lesion origin as melanocytic or non-melanocytic with an accuracy of 87%, predict lesion malignancy with 77% accuracy, and provide disease diagnosis with an accuracy of 71%. In addition, the implemented interpretability methods assist understand the model's decision-making process and detecting misdiagnoses. This work is a step toward achieving interpretability in skin cancer diagnosis using clinical images. The developed approach can assist general practitioners to make an early diagnosis, thus reducing the redundant referrals that expert dermatologists receive for further investigations.

A comparison of manual and automated neural architecture search for white matter tract segmentation

Segmentation of white matter tracts in diffusion magnetic resonance images is an important first step in many imaging studies of the brain in health and disease. Similar to medical image segmentation in general, a popular approach to white matter tract segmentation is to use U-Net based artificial neural network architectures. Despite many suggested improvements to the U-Net architecture in recent years, there is a lack of systematic comparison of architectural variants for white matter tract segmentation. In this paper, we evaluate multiple U-Net based architectures specifically for this purpose. We compare the results of these networks to those achieved by our own various architecture changes, as well as to new U-Net architectures designed automatically via neural architecture search (NAS). To the best of our knowledge, this is the first study to systematically compare multiple U-Net based architectures for white matter tract segmentation, and the first to use NAS. We find that the recently proposed medical imaging segmentation network UNet3+ slightly outperforms the current state of the art for white matter tract segmentation, and achieves a notably better mean Dice score for segmentation of the fornix (+ 0.01 and + 0.006 mean Dice increase for left and right fornix respectively), a tract that the current state of the art model struggles to segment. UNet3+ also outperforms the current state of the art when little training data is available. Additionally, manual architecture search found that a minor segmentation improvement is observed when an additional, deeper layer is added to the U-shape of UNet3+. However, all networks, including those designed via NAS, achieve similar results, suggesting that there may be benefit in exploring networks that deviate from the general U-Net paradigm.

A deep learning approach for semantic segmentation of unbalanced data in electron tomography of catalytic materials

In computed TEM tomography, image segmentation represents one of the most basic tasks with implications not only for 3D volume visualization, but more importantly for quantitative 3D analysis. In case of large and complex 3D data sets, segmentation can be an extremely difficult and laborious task, and thus has been one of the biggest hurdles for comprehensive 3D analysis. Heterogeneous catalysts have complex surface and bulk structures, and often sparse distribution of catalytic particles with relatively poor intrinsic contrast, which possess a unique challenge for image segmentation, including the current state-of-the-art deep learning methods. To tackle this problem, we apply a deep learning-based approach for the multi-class semantic segmentation of a γ-Alumina/Pt catalytic material in a class imbalance situation. Specifically, we used the weighted focal loss as a loss function and attached it to the U-Net's fully convolutional network architecture. We assessed the accuracy of our results using Dice similarity coefficient (DSC), recall, precision, and Hausdorff distance (HD) metrics on the overlap between the ground-truth and predicted segmentations. Our adopted U-Net model with the weighted focal loss function achieved an average DSC score of 0.96 ± 0.003 in the γ-Alumina support material and 0.84 ± 0.03 in the Pt NPs segmentation tasks. We report an average boundary-overlap error of less than 2 nm at the 90th percentile of HD for γ-Alumina and Pt NPs segmentations. The complex surface morphology of γ-Alumina and its relation to the Pt NPs were visualized in 3D by the deep learning-assisted automatic segmentation of a large data set of high-angle annular dark-field (HAADF) scanning transmission electron microscopy (STEM) tomography reconstructions.

Deep morphological recognition of kidney stones using intra-operative endoscopic digital videos

Objective. To assess the performance and added value of processing complete digital endoscopic video sequences for the automatic recognition of stone morphological features during a standard-of-care intra-operative session.Approach. A computer-aided video classifier was developed to predict in-situ the morphology of stone using an intra-operative digital endoscopic video acquired in a clinical setting. Using dedicated artificial intelligence (AI) networks, the proposed pipeline selects adequate frames in steady sequences of the video, ensures the presence of (potentially fragmented) stones and predicts the stone morphologies on a frame-by-frame basis. The automatic endoscopic stone recognition (A-ESR) is subsequently carried out by mixing all collected morphological observations.Main results. The proposed technique was evaluated on pure (i.e. include one morphology) and mixed (i.e. include at least two morphologies) stones involving ‘Ia/Calcium Oxalate Monohydrate’ (COM), ‘IIb/Calcium Oxalate Dihydrate’ (COD) and ‘IIIb/Uric Acid’ (UA) morphologies. The gold standard ESR was provided by a trained endo-urologist and confirmed by microscopy and infra-red spectroscopy. For the AI-training, 585 static images were collected (349 and 236 observations of stone surface and section, respectively) and used. Using the proposed video classifier, 71 digital endoscopic videos were analyzed: 50 exhibited only one morphological type and 21 displayed two. Taken together, both pure and mixed stone types yielded a mean diagnostic performances as follows: balanced accuracy = [88 ± 6] (min = 81)%, sensitivity = [80 ± 13] (min = 69)%, specificity = [95 ± 2] (min = 92)%, precision = [78 ± 12] (min = 62)% and F1-score = [78 ± 7] (min = 69)%.Significance. These results demonstrate that AI applied on digital endoscopic video sequences is a promising tool for collecting morphological information during the time-course of the stone fragmentation process without resorting to any human intervention for stone delineation or the selection of adequate steady frames.

Deep 3D Neural Network for Brain Structures Segmentation Using Self-Attention Modules in MRI Images

In recent years, the use of deep learning-based models for developing advanced healthcare systems has been growing due to the results they can achieve. However, the majority of the proposed deep learning-models largely use convolutional and pooling operations, causing a loss in valuable data and focusing on local information. In this paper, we propose a deep learning-based approach that uses global and local features which are of importance in the medical image segmentation process. In order to train the architecture, we used extracted three-dimensional (3D) blocks from the full magnetic resonance image resolution, which were sent through a set of successive convolutional neural network (CNN) layers free of pooling operations to extract local information. Later, we sent the resulting feature maps to successive layers of self-attention modules to obtain the global context, whose output was later dispatched to the decoder pipeline composed mostly of upsampling layers. The model was trained using the Mindboggle-101 dataset. The experimental results showed that the self-attention modules allow segmentation with a higher Mean Dice Score of 0.90 ± 0.036 compared with other UNet-based approaches. The average segmentation time was approximately 0.038 s per brain structure. The proposed model allows tackling the brain structure segmentation task properly. Exploiting the global context that the self-attention modules incorporate allows for more precise and faster segmentation. We segmented 37 brain structures and, to the best of our knowledge, it is the largest number of structures under a 3D approach using attention mechanisms.

Deep Learning and Domain-Specific Knowledge to Segment the Liver from Synthetic Dual Energy CT Iodine Scans

We map single energy CT (SECT) scans to synthetic dual-energy CT (synth-DECT) material density iodine (MDI) scans using deep learning (DL) and demonstrate their value for liver segmentation. A 2D pix2pix (P2P) network was trained on 100 abdominal DECT scans to infer synth-DECT MDI scans from SECT scans. The source and target domain were paired with DECT monochromatic 70 keV and MDI scans. The trained P2P algorithm then transformed 140 public SECT scans to synth-DECT scans. We split 131 scans into 60% train, 20% tune, and 20% held-out test to train four existing liver segmentation frameworks. The remaining nine low-dose SECT scans tested system generalization. Segmentation accuracy was measured with the dice coefficient (DSC). The DSC per slice was computed to identify sources of error. With synth-DECT (and SECT) scans, an average DSC score of 0.93±0.06 (0.89±0.01) and 0.89±0.01 (0.81±0.02) was achieved on the held-out and generalization test sets. Synth-DECT-trained systems required less data to perform as well as SECT-trained systems. Low DSC scores were primarily observed around the scan margin or due to non-liver tissue or distortions within ground-truth annotations. In general, training with synth-DECT scans resulted in improved segmentation performance with less data.

Deep structured residual encoder-decoder network with a novel loss function for nuclei segmentation of kidney and breast histopathology images

To improve the process of diagnosis and treatment of cancer disease, automatic segmentation of haematoxylin and eosin (H & E) stained cell nuclei from histopathology images is the first step in digital pathology. The proposed deep structured residual encoder-decoder network (DSREDN) focuses on two aspects: first, it effectively utilized residual connections throughout the network and provides a wide and deep encoder-decoder path, which results to capture relevant context and more localized features. Second, vanished boundary of detected nuclei is addressed by proposing an efficient loss function that better train our proposed model and reduces the false prediction which is undesirable especially in healthcare applications. The proposed architecture experimented on three different publicly available H&E stained histopathological datasets namely: (I) Kidney (RCC) (II) Triple Negative Breast Cancer (TNBC) (III) MoNuSeg-2018. We have considered F1-score, Aggregated Jaccard Index (AJI), the total number of parameters, and FLOPs (Floating point operations), which are mostly preferred performance measure metrics for comparison of nuclei segmentation. The evaluated score of nuclei segmentation indicated that the proposed architecture achieved a considerable margin over five state-of-the-art deep learning models on three different histopathology datasets. Visual segmentation results show that the proposed DSREDN model accurately segment the nuclear regions than those of the state-of-the-art methods.

Deep Learning for the Automatic Quantification of Pleural Plaques in Asbestos-Exposed Subjects

Objective: This study aimed to develop and validate an automated artificial intelligence (AI)-driven quantification of pleural plaques in a population of retired workers previously occupationally exposed to asbestos. Methods: CT scans of former workers previously occupationally exposed to asbestos who participated in the multicenter APEXS (Asbestos PostExposure Survey) study were collected retrospectively between 2010 and 2017 during the second and the third rounds of the survey. A hundred and forty-one participants with pleural plaques identified by expert radiologists at the 2nd and the 3rd CT screenings were included. Maximum Intensity Projection (MIP) with 5 mm thickness was used to reduce the number of CT slices for manual delineation. A Deep Learning AI algorithm using 2D-convolutional neural networks was trained with 8280 images from 138 CT scans of 69 participants for the semantic labeling of Pleural Plaques (PP). In all, 2160 CT images from 36 CT scans of 18 participants were used for AI testing versus ground-truth labels (GT). The clinical validity of the method was evaluated longitudinally in 54 participants with pleural plaques. Results: The concordance correlation coefficient (CCC) between AI-driven and GT was almost perfect (>0.98) for the volume extent of both PP and calcified PP. The 2D pixel similarity overlap of AI versus GT was good (DICE = 0.63) for PP, whether they were calcified or not, and very good (DICE = 0.82) for calcified PP. A longitudinal comparison of the volumetric extent of PP showed a significant increase in PP volumes (p < 0.001) between the 2nd and the 3rd CT screenings with an average delay of 5 years. Conclusions: AI allows a fully automated volumetric quantification of pleural plaques showing volumetric progression of PP over a five-year period. The reproducible PP volume evaluation may enable further investigations for the comprehension of the unclear relationships between pleural plaques and both respiratory function and occurrence of thoracic malignancy.

Deciphering tumour tissue organization by 3D electron microscopy and machine learning

Despite recent progress in the characterization of tumour components, the tri-dimensional (3D) organization of this pathological tissue and the parameters determining its internal architecture remain elusive. Here, we analysed the spatial organization of patient-derived xenograft tissues generated from hepatoblastoma, the most frequent childhood liver tumour, by serial block-face scanning electron microscopy using an integrated workflow combining 3D imaging, manual and machine learning-based semi-automatic segmentations, mathematics and infographics. By digitally reconstituting an entire hepatoblastoma sample with a blood capillary, a bile canaliculus-like structure, hundreds of tumour cells and their main organelles (e.g. cytoplasm, nucleus, mitochondria), we report unique 3D ultrastructural data about the organization of tumour tissue. We found that the size of hepatoblastoma cells correlates with the size of their nucleus, cytoplasm and mitochondrial mass. We also found anatomical connections between the blood capillary and the planar alignment and size of tumour cells in their 3D milieu. Finally, a set of tumour cells polarized in the direction of a hot spot corresponding to a bile canaliculus-like structure. In conclusion, this pilot study allowed the identification of bioarchitectural parameters that shape the internal and spatial organization of tumours, thus paving the way for future investigations in the emerging onconanotomy field.

de Senneville et al. demonstrate an integrated workflow combining 3D imaging, manual and machine learning-based semi-automatic segmentation, mathematics and infographics to study the spatial organization of patient-derived hepatoblastoma xenograft tissues. Their approach potentially assists investigations of this childhood liver tumour and other types of tumour tissues.