Technical note: Progressive deep learning: An accelerated training strategy for medical image segmentation.

MP43-02 LESSONS LEARNED IN APPLYING DEEP LEARNING TO FACILITATE PROSTATE MR-US FUSION BIOPSY WORKFLOW

Segmentation of breast ultrasound images is a crucial and challenging task in computer-aided diagnosis systems. Accurately segmenting masses in benign and malignant cases and identifying regions with no mass is a primary objective in breast ultrasound image segmentation. Deep learning (DL) has emerged as a powerful tool in medical image segmentation, revolutionizing how medical professionals analyze and interpret complex imaging data. The UNet architecture is a highly regarded and widely used DL model in medical image segmentation. Its distinctive architectural design and exceptional performance have made it popular among researchers. With the increase in data and model complexity, optimization and fine-tuning models play a vital and more challenging role than before. This paper presents a comparative study evaluating the effect of image preprocessing and different optimization techniques and the importance of fine-tuning different UNet segmentation models for breast ultrasound images. Optimization and fine-tuning techniques have been applied to enhance the performance of UNet, Sharp UNet, and Attention UNet. Building upon this progress, we designed a novel approach by combining Sharp UNet and Attention UNet, known as Sharp Attention UNet. Our analysis yielded the following quantitative evaluation metrics for the Sharp Attention UNet: the Dice coefficient, specificity, sensitivity, and F1 score values obtained were 0.93, 0.99, 0.94, and 0.94, respectively. In addition, McNemar’s statistical test was applied to assess significant differences between the approaches. Across a number of measures, our proposed model outperformed all other models, resulting in improved breast lesion segmentation.

Introduction: Magnetic resonance imaging (MRI) is the most common tool to examine glioblastoma. Preoperative MRI can be used to initial diagnosis, and surgery planning, while follow-up MRIs can be used to evaluate treatment responses, identify recurrency, and detect side effects. The follow-up MRIs are usually taken after the first-line therapy, such as maximal safe resection. In the past, radiologists manually segment the tumor regions from normal brain tissue on follow-up MRIs, which is time-consuming, error-prone, and challenging. Several deep-learning (DL) models have been developed utilizing preoperative images, but their performance has yet to be evaluated on follow-up MRIs. In this research, we built the largest follow-up MRI cohort (311 patients) to assess these DL models and their generalizability and performance on independent preoperative and follow-up images. We also made the first follow-up-based deep learning models for this specific task. Methods: All evaluation deep learning models (10 models) were trained by the Brain Tumor Segmentation challenge 2020 (BraTS’20) and evaluated by fifty pairs of preoperative and follow-up scans from our institution. The segmentation form our institution is evaluated by board certified radiologist. MRIs in the BraTS’20 dataset were all preoperative scans. After the evaluation, we randomly assigned 264 patients’ scans from our institution to the training dataset and 47 patients’ scans to the testing dataset. We compared three types of models in our follow-up deep learning model, including 1) UNet-3D, 2) UNet-3D+transfer-learning, and 3) UNet-3D+transfer-learning+baysian-learning. The benchmark for all models was the Dice similarity coefficient (DSC). DSC can measure the spatial overlap between model prediction and ground truth. The value of DSC is between zero to one. Zero means no overlap and one indicates complete overlap. Results: Our study demonstrates that the BraTS'20 trained models' performance decreased by 13.05% in independent preoperative MRI scans and 19.04% in follow-up MRI scans. The most significant mismatch regions were FLAIR hyperintense regions (3.68% drop in independent preoperative scans and 10.61% drop in independent follow-up scans) and Non-enhancing core (5.20% drop in independent preoperative scans and 11.99% drop in independent follow-up scans). Our best model can achieve the best DSC among three tumor regions compared to all evaluation models (FLAIR hyperintense regions: DSC 0.77 V.S. DSC 0.58; Enhancing tumor region: DSC 0.87 V.S. DSC 0.68; Non-enhancing tumor region: DSC 0.92 V.S. DSC 0.55). Conclusion: Maximal safe resection induced brain structure change, decreasing the performance of the preoperative-based DL model. Implementing a follow-up MRI-based segmentation model is essential to make accurate and generalizable results to address structural changes after maximal safe resection. Our follow-up DL model demonstrates the DSC score can be recovered. We commit to further developing the tool to assist radiologists in handling follow-up MRIs for glioblastoma patients. Citation Format: Kang Lin Hsieh, Tanjida Kabir, Luis Nunez-Rubiano, Yu-Chun Hsu, Yu Cai, Juan Rodriguez Quintero, Octavio Arevalo, Kangyi Zhao, Jackie Zhang, Jiguang Zhu Zhu, Roy Riascos, Xioaqian Jiang, Shayan Shams. A confident and operator-independent deep segmentation model to measure residual tumor volume in the follow-up MRIs for glioblastoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 2 (Clinical Trials and Late-Breaking Research); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(8_Suppl):Abstract nr LB067.

With the increasing integration of artificial intelligence (AI) in several scientific domains, there is a rising demand for advanced AI tools capable of addressing advanced research challenges. A challenge of paramount importance lies in accurately predicting the streamflow within river basins. Effective river flow prediction holds significant relevance, particularly given the substantial societal implications of river usage, encompassing areas such as transportation, agriculture, and power generation. The present study introduces a novel approach to streamflow prediction involving the development of a Deep Learning (DL) model that combines a convolutional neural network with Transfer Learning (TL) techniques to predict streamflow in river systems. With the aim of training the developed DL model, the study employed a time-series dataset containing hydrological data related to two distinct river basins, i.e., Paraíba do Sul, in Brazil, and Zambezi in the state of Mozambique. The developed DL models exhibited the capability to effectively predict the river flow with a one-day horizon, relying on the preceding three or seven days of historical data. To overcome the limited availability of training data and reduce the training time of DL models, TL was leveraged to incorporate two additional distinct time-series datasets, i.e., historical streamflow data from the São Francisco River in Brazil, and climate data from Delhi, India. The application of TL significantly reduced training time, leading only to a minimal decrease in prediction performance. Indeed, in the case DL models were trained on data collected from the Paraíba do Sul River, a substantial reduction in training time was observed - up to 27% - with a modest percentage decrease of 0.31% in test predictive performance (). Similarly, TL induced a significant reduction in training time of up to 48%, while resulting in a modest 2% reduction in test predictive performance () for the Zambezi dataset. The findings underscore the significance of TL as a strategic and viable approach to improve the efficiency of river flow prediction models in the context of basins with limited hydrological data available.

Research in many fields has shown that transfer learning (TL) is well-suited to improve the performance of deep learning (DL) models in datasets with small numbers of samples. This empirical success has triggered interest in the application of TL to cognitive decoding analyses with functional neuroimaging data. Here, we systematically evaluate TL for the application of DL models to the decoding of cognitive states (e.g., viewing images of faces or houses) from whole-brain functional Magnetic Resonance Imaging (fMRI) data. We first pre-train two DL architectures on a large, public fMRI dataset and subsequently evaluate their performance in an independent experimental task and a fully independent dataset. The pre-trained DL models consistently achieve higher decoding accuracies and generally require less training time and data than model variants that were not pre-trained, while also outperforming linear baseline models trained from scratch, clearly underlining the benefits of pre-training. We demonstrate that these benefits arise from the ability of the pre-trained models to reuse many of their learned features when training with new data, providing deeper insights into the mechanisms giving rise to the benefits of pre-training. Yet, we also surface nuanced challenges for whole-brain cognitive decoding with DL models when interpreting the decoding decisions of the pre-trained models, as these have learned to utilize the fMRI data in unforeseen and counterintuitive ways to identify individual cognitive states.

Aim To develop an optimal method for automated segmentation of atherosclerotic plaque structural components in optical coherence tomography (OCT) images using an ensemble of deep learning neural network models based on a comparison of nine artificial neural network architectures.Material and methods This study utilized a multidisciplinary OCT dataset obtained from examinations of 103 patients. Each OCT session was annotated as a set of two-dimensional binary masks corresponding to the pixel boundaries of four key morphological features of plaques: vascular lumen, fibrous cap, lipid core, and microvessels. Nine deep learning models, including U-Net, DeepLabV3, and others, were used to segment the anatomical structures. Model hyperparameters were tuned using Bayesian optimization, and the performance was assessed using the Dice Similarity Coefficient (DSC) metric.Results The models demonstrated high segmentation accuracy for the vascular lumen (DSC: 0.987) and moderate accuracy for the fibrous cap (DSC: 0.736) and lipid core (DSC: 0.751). Microvessel segmentation proved more challenging, with a final DSC accuracy of 61%. A weighted ensemble of models, accounting for the uneven representation of different morphological structures and model confidence, achieved an average DSC of 88.2%, which significantly improved the overall accuracy compared to individual models. This accuracy value exceeds those of all individual models, including the highest DSC values of 0.784 (for microvessels) and 0.751 (for lipid core), indicating a consistent improvement in the segmentation accuracy by integrating the models.Conclusion The study confirms the effectiveness of the proposed strategy for segmentation of atherosclerotic plaques in OCT images. This strategy is based on using specialized models for various morphological features and a weighted ensemble adapted to the uneven representation of different morphological structures and morphological complexity. This approach provided a high segmentation accuracy despite the pronounced inequality in the representation of classes. The study results can contribute to the development of decision support methods in cardiology aimed at improving diagnostic accuracy and treatment of cardiovascular diseases.

Implementation of automation routines leveraging deep learning (DL) methods has been a growing topic of interest. The focus of this work is on the Gamma Knife (GK) workflow, specifically the approximation of dose distributions from GK plan parameters and the prediction of dose distributions using DL. These works contribute towards the larger goal of treatment plan prediction and are seen as intermediate steps towards that end. The approximation of dose distributions was motivated by the closed nature of the GK system, which causes complications with the evaluation of the predictions made by the DL models. The approximation utilizes a superposition method, accumulating dose from a basis set of dose distributions which correspond to specific sector activations of the GK system. The approximation method was compared to the original dose distributions calculated by the GK treatment planning system (TPS), Leksell GammaPlan, for 30 patients using the similarity metrics of Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD) as well as the plan quality metrics of coverage, selectivity, and gradient index. In the range of clinical planning significance, the metrics reported strong agreement to the original calculation (DSC > 0.85, HD < 2.5mm, coverage difference = 0.014 ± 0.035, selectivity difference = 0.008 ± 0.037, and gradient index difference = 0.118 ± 0.192). The second focus of this work was to improve the performance of DL models used to predict dose distributions from target contour information. This was accomplished by applying transforms to the binary target contour volumes, resulting in alternate model input channels that encode the distance of each voxel to the boundary of the nearest contour, both within and outside of the contours. Training and evaluation were performed with a standard UNet and HD-UNet architecture for large and small targets separately. DSC and Predictions resulting from these encodings resulted in significantly better predictions (p < 0.05) than their binary target counterparts at all (excluding 90% isodose for the HD-Unet) threshold levels for small targets. The large target trainings were significantly (p < 0.05) improved in the mid-to-high dose ranges.

Deep learning and convolutional neural networks (CNNs) in particular are increasingly popular tools for segmentation and classification of medical images. CNNs were shown to be successful for segmentation of brain tumors into multiple regions or labels. However, in the environment which fosters data-sharing and collection of multi-institutional datasets, a question arises: does training with data from another institution with potentially different imaging equipment, contrast protocol, and patient population impact the segmentation performance of the CNN? Our study presents preliminary data towards answering this question. Specifically, we used MRI data of glioblastoma (GBM) patients for two institutions present in The Cancer Imaging Archive. We performed a process of training and testing CNN multiple times such that half of the time the CNN was tested on data from the same institution that was used for training and half of the time it was tested on another institution, keeping the training and testing set size constant. We observed a decrease in performance as measured by Dice coefficient when the CNN was trained with data from a different institution as compared to training with data from the same institution. The changes in performance for the entire tumor and for four different labels within the tumor were: 0.72 to 0.65 (p=0.06), 0.61 to 0.58 (p=0.49), 0.54 to 0.51 (p=0.82), 0.31 to 0.24 (p<0.03), and 0.43 to 0.31(p<0.003) respectively. In summary, we found that while data across institutions can be used for development of CNNs, this might be associated with a decrease in performance.

Detecting programmed death ligand 1 (PD-L1) expression based on immunohistochemical (IHC) staining is an important guide for the treatment of lung cancer with immune checkpoint inhibitors. However, this method has problems such as high staining costs, tumor heterogeneity, and subjective differences among pathologists. Therefore, the application of deep learning models to segment and quantitatively predict PD-L1 expression in digital sections of Hematoxylin and eosin (H&E) stained lung squamous cell carcinoma is of great significance. We constructed a dataset comprising H&E-stained digital sections of lung squamous cell carcinoma and used a Transformer Unet (TransUnet) deep learning network with an encoder-decoder design to segment PD-L1 negative and positive regions and quantitatively predict the tumor cell positive score (TPS). The results showed that the dice similarity coefficient (DSC) and intersection overunion (IoU) of deep learning for PD-L1 expression segmentation of H&E-stained digital slides of lung squamous cell carcinoma were 80 and 72%, respectively, which were better than the other seven cutting-edge segmentation models. The root mean square error (RMSE) of quantitative prediction TPS was 26.8, and the intra-group correlation coefficients with the gold standard was 0.92 (95% CI: 0.90-0.93), which was better than the consistency between the results of five pathologists and the gold standard. The deep learning model is capable of segmenting and quantitatively predicting PD-L1 expression in H&E-stained digital sections of lung squamous cell carcinoma, which has significant implications for the application and guidance of immune checkpoint inhibitor treatments. And the link to the code is https://github.com/Baron-Huang/PD-L1-prediction-via-HE-image.

This research paper explores the rising interest in utilising deep learning methods, specifically Convolution Neural Network (CNN), to improve the reconstruction of Compressively Sampled Magnetic Resonance Imaging (CS MRI) images from under-sampled data. By training deep learning architectures on extensive data sets of paired under sampled and fully sampled MR images, these models aim to capture intricate patterns and structures, ultimately enhancing the accuracy of MR image reconstructions. A novel deep learning model is proposed dubbed as “Deep Iterative Shrinkage Thresholding Algorithm-Compressed Sensing Network” (DISTA-CSNet), specifically designed for efficient recovery of CS MRI. Our model showcases impressive results with only 20 epochs and can be effortlessly trained on diverse datasets. To ensure robustness across different datasets, the dropouts are incorporated into the model, evident from testing results. The trained DISTA-CSNet exhibits remarkable performance in recovering CS MRI from various data sets, surpassing several advanced deep learning techniques with changing compression ratios consistently. The proposed model demonstrates significant enhancements in both Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) metrics, affirming the efficacy of our proposed model. The capability of our DISTA-CSNet in accurately reconstructing CS MRI images from 5-fold undersampled data shows promise in improving medical imaging applications and advancing the field of compressed sensing MRI.

Accurate prostate volume estimation is crucial for effective prostate disease management. Ultrasound (US) imaging, particularly transrectal ultrasound, offers a cost-effective and rapid assessment. However, US images often suffer from artifacts and poor contrast, making prostate volume estimation challenging. This review explores recent advancements in deep learning (DL) techniques for automatic prostate segmentation in US images as a primary step toward prostate volume estimation. We examine various DL architectures, including traditional U-Net modifications and innovative designs incorporating residual connections, multi-directional image data, and attention mechanisms. Additionally, we discuss pre-processing methods to enhance image quality, the integration of shape information, and strategies to improve the consistency and robustness of DL models. The effectiveness of these techniques is evaluated through metrics such as the Dice Similarity Coefficient, Jaccard Index, and Hausdorff Distance. The review highlights the potential of DL in improving prostate volume estimation accuracy and reducing clinical workload while also identifying areas for future research to enhance model performance and generalizability.

Monte Carlo (MC)-based estimation of dose-averaged linear energy transfer (LETd) in proton therapy is accurate but computationally intensive. Although deep learning (DL) models offer efficient alternatives, few have been optimized for multi-source inputs or validated on anatomically complex tumors, such as nasopharyngeal carcinoma (NPC). This study developed and evaluated a multi-source DL framework for voxel-level LETd prediction in proton therapy for NPC, aiming to identify optimal input configurations and assess both prediction accuracy and clinical reliability. This study analyzed 100 patients with NPC, yielding 500 proton therapy fields. Ground-truth LETd maps were generated using GPU-based Monte Carlo simulations. A deep learning model was trained with different combinations of four input features: dose distribution, CT image, beam mask, and radiological depth, resulting in eight configurations (Mdose, Mdose+ct, Mdose+beam, Mdose+depth, Mdose+beam+depth, Mdose+ct+depth, Mdose+ct+beam, and Mdose+ct+beam+depth). The dataset was randomly divided into training (300 fields, 60 patients), validation (100 fields, 20 patients), and testing (100 fields, 20 patients) sets. Model performance was evaluated using mean absolute error (MAE), peak signal-to-noise ratio (PSNR), normalized cross-correlation (NCC), structural similarity index measure (SSIM), and 3D Gamma analysis. A 3D U-Net baseline was retrained under identical conditions for comparison. Model interpretability was examined with Gradient-weighted Class Activation Mapping (Grad-CAM), and prediction confidence was assessed using uncertainty maps. The Mdose+ct model achieved the best performance (MAE 0.21±0.03keV/µm, PSNR 28.99±1.33, NCC 0.98±0.01, SSIM 0.971±0.007, Gamma 3%/3mm: 94.69±3.69%), outperforming the 3D U-Net baseline (MAE 0.38±0.52keV/µm). Grad-CAM visualizations confirmed that the model focused on physically relevant regions, while uncertainty maps highlighted high-variability zones at LETd transitions. The model generated voxel-level LETd predictions in approximately 7 s per case, which was substantially faster than GPU-based MC simulations (30 s per case). The proposed multi-source deep learning framework enables rapid and accurate voxel-level LETd prediction for NPC. Combining dose with CT yielded the most reliable input configuration, and incorporating interpretability and uncertainty analyses enhances confidence in its suitability for biologically guided proton therapy planning.

In recent days, deep learning is on rage and is gaining a huge amount of popularity due to its supremacy in terms of accuracy. Deep learning is being used for a vast number of applications out of which healthcare is an important category. In this paper, we discuss the role of deep learning in medical image segmentation. It is also known as the automated or semi-automated detection of edges within various medical image modalities so as to identify the region of interest. Furthermore, we also explore the various deep learning networks that are widely preferred for medical image segmentation along with the architecture and overview of each network. This paper covers the most recent and widely preferred deep learning networks such as Convolutional Neural Network (CNN) and other related networks such as Alexnet, Resnet, U-net and V-net. The challenges and limitations of the emerging DL networks is also studied.KeywordsDeep learningImage segmentationCNNAlexnetResnetU-netV-net

Early detection and assessment of polyps play a crucial role in the prevention and treatment of colorectal cancer (CRC). Polyp segmentation provides an effective solution to assist clinicians in accurately locating and segmenting polyp regions. In the past, people often relied on manually extracted lower-level features such as color, texture, and shape, which often had problems capturing global context and lacked robustness to complex scenarios. With the advent of deep learning, more and more medical image segmentation algorithms based on deep learning networks have emerged, making significant progress in the field. This paper provides a comprehensive review of polyp segmentation algorithms. We first review some traditional algorithms based on manually extracted features and deep segmentation algorithms, and then describe benchmark datasets related to the topic. Specifically, we carry out a comprehensive evaluation of recent deep learning models and results based on polyp size, taking into account the focus of research topics and differences in network structures. Finally, we discuss the challenges of polyp segmentation and future trends in the field.

Image segmentation is the process of delineating the regions occupied by objects of interest in a given image. This operation is a fundamentally required first step in numerous applications of medical imagery. In the medical imaging field, this activity has a rich literature that spans over 45 years. In spite of numerous advances, including deep learning (DL) networks (DLNs) in recent years, the problem has defied a robust, fail-safe, and satisfactory solution, especially for objects that are manifest with low contrast, are spatially sparse, have variable shape among individuals, or are affected by imaging artifacts, pathology, or post-treatment change in the body. Although image processing techniques, notably DLNs, are uncanny in their ability to harness low-level intensity pattern information on objects, they fall short in the high-level task of identifying and localizing an entire object as a gestalt. This dilemma has been a fundamental unmet challenge in medical image segmentation. In this paper, we demonstrate that by synergistically marrying the unmatched strengths of high-level human knowledge (i.e., natural intelligence (NI)) with the capabilities of DL networks (i.e., artificial intelligence (AI)) in garnering intricate details, these challenges can be significantly overcome. Focusing on the object recognition task, we formulate an anatomy-guided DL object recognition approach named Automatic Anatomy Recognition-Deep Learning (AAR-DL) which combines an advanced anatomy-modeling strategy, model-based non-DL object recognition, and DL object detection networks to achieve expert human-like performance.