Epiretinal Membrane Detection at the Ophthalmologist Level using Deep Learning of Optical Coherence Tomography

PURPOSE

Previous deep learning studies on optical coherence tomography (OCT) mainly focused on diabetic retinopathy and age-related macular degeneration. We proposed a deep learning model that can identify epiretinal membrane (ERM) in OCT with ophthalmologist-level performance.

DESIGN

Cross-sectional study.

PARTICIPANTS

A total of 3,618 central fovea cross section OCT images from 1,475 eyes of 964 patients.

METHODS

We retrospectively collected 7,652 OCT images from 1,197 patients. From these images, 2,171 were normal and 1,447 were ERM OCT. A total of 3,141 OCT images was used as training dataset and 477 images as testing dataset. DL algorithm was used to train the interpretation model. Diagnostic results by four board-certified non-retinal specialized ophthalmologists on the testing dataset were compared with those generated by the DL model.

MAIN OUTCOME MEASURES

We calculated for the derived DL model the following characteristics: sensitivity, specificity, F1 score and area under curve (AUC) of the receiver operating characteristic (ROC) curve. These were calculated according to the gold standard results which were parallel diagnoses of the retinal specialist. Performance of the DL model was finally compared with that of non-retinal specialized ophthalmologists.

RESULTS

Regarding the diagnosis of ERM in OCT images, the trained DL model had the following characteristics in performance: sensitivity: 98.7%, specificity: 98.0%, and F1 score: 0.945. The accuracy on the training dataset was 99.7% (95% CI: 99.4 - 99.9%), and for the testing dataset, diagnostic accuracy was 98.1% (95% CI: 96.5 - 99.1%). AUC of the ROC curve was 0.999. The DL model slightly outperformed the average non-retinal specialized ophthalmologists.

CONCLUSIONS

An ophthalmologist-level DL model was built here to accurately identify ERM in OCT images. The performance of the model was slightly better than the average non-retinal specialized ophthalmologists. The derived model may play a role to assist clinicians to promote the efficiency and safety of healthcare in the future.

Implementation of a parallel protein structure alignment service on cloud

Protein structure alignment has become an important strategy by which to identify evolutionary relationships between protein sequences. Several alignment tools are currently available for online comparison of protein structures. In this paper, we propose a parallel protein structure alignment service based on the Hadoop distribution framework. This service includes a protein structure alignment algorithm, a refinement algorithm, and a MapReduce programming model. The refinement algorithm refines the result of alignment. To process vast numbers of protein structures in parallel, the alignment and refinement algorithms are implemented using MapReduce. We analyzed and compared the structure alignments produced by different methods using a dataset randomly selected from the PDB database. The experimental results verify that the proposed algorithm refines the resulting alignments more accurately than existing algorithms. Meanwhile, the computational performance of the proposed service is proportional to the number of processors used in our cloud platform.

GPU-Accelerated GLRLM Algorithm for Feature Extraction of MRI

The gray level run length matrix (GLRLM) whose entries are statistics recording distribution and relationship of images pixels is a widely used method for extracting statistical features for medical images, e.g., magnetic resonance (MR) images. Recently these features are usually employed in some artificial neural networks to identify and distinguish texture patterns. But GLRLM construction and features extraction are tedious and computationally intensive while the images are too big with high resolution, or there are too many small or intermediate Regions of Interest (ROI) to process in a single image, which makes the preprocess a time consuming stage. Hence, it is of great importance to accelerate the procedure which is nowadays possible with the rapid development of massively parallel Graphics Processing Unit, i.e. the GPU computing technology. In this article, we propose a new paradigm based on mature parallel primitives for generating GLRLMs and extracting multiple features for many ROIs simultaneously in a single image. Experiments show that such a paradigm is easy to implement and offers an acceleration over 5 fold increase in speed than an optimized serial counterpart.

Cloud computing for protein-ligand binding site comparison

The proteome-wide analysis of protein-ligand binding sites and their interactions with ligands is important in structure-based drug design and in understanding ligand cross reactivity and toxicity. The well-known and commonly used software, SMAP, has been designed for 3D ligand binding site comparison and similarity searching of a structural proteome. SMAP can also predict drug side effects and reassign existing drugs to new indications. However, the computing scale of SMAP is limited. We have developed a high availability, high performance system that expands the comparison scale of SMAP. This cloud computing service, called Cloud-PLBS, combines the SMAP and Hadoop frameworks and is deployed on a virtual cloud computing platform. To handle the vast amount of experimental data on protein-ligand binding site pairs, Cloud-PLBS exploits the MapReduce paradigm as a management and parallelizing tool. Cloud-PLBS provides a web portal and scalability through which biologists can address a wide range of computer-intensive questions in biology and drug discovery.

Local alignment tool based on Hadoop framework and GPU architecture

With the rapid growth of next generation sequencing technologies, such as Slex, more and more data have been discovered and published. To analyze such huge data the computational performance is an important issue. Recently, many tools, such as SOAP, have been implemented on Hadoop and GPU parallel computing architectures. BLASTP is an important tool, implemented on GPU architectures, for biologists to compare protein sequences. To deal with the big biology data, it is hard to rely on single GPU. Therefore, we implement a distributed BLASTP by combining Hadoop and multi-GPUs. The experimental results present that the proposed method can improve the performance of BLASTP on single GPU, and also it can achieve high availability and fault tolerance.

Efficient haplotype block partitioning and tag SNP selection algorithms under various constraints

Patterns of linkage disequilibrium plays a central role in genome-wide association studies aimed at identifying genetic variation responsible for common human diseases. These patterns in human chromosomes show a block-like structure, and regions of high linkage disequilibrium are called haplotype blocks. A small subset of SNPs, called tag SNPs, is sufficient to capture the haplotype patterns in each haplotype block. Previously developed algorithms completely partition a haplotype sample into blocks while attempting to minimize the number of tag SNPs. However, when resource limitations prevent genotyping all the tag SNPs, it is desirable to restrict their number. We propose two dynamic programming algorithms, incorporating many diversity evaluation functions, for haplotype block partitioning using a limited number of tag SNPs. We use the proposed algorithms to partition the chromosome 21 haplotype data. When the sample is fully partitioned into blocks by our algorithms, the 2,266 blocks and 3,260 tag SNPs are fewer than those identified by previous studies. We also demonstrate that our algorithms find the optimal solution by exploiting the nonmonotonic property of a common haplotype-evaluation function.

Accelerating Smith-Waterman Alignment for Protein Database Search Using Frequency Distance Filtration Scheme Based on CPU-GPU Collaborative System

The Smith-Waterman (SW) algorithm has been widely utilized for searching biological sequence databases in bioinformatics. Recently, several works have adopted the graphic card with Graphic Processing Units (GPUs) and their associated CUDA model to enhance the performance of SW computations. However, these works mainly focused on the protein database search by using the intertask parallelization technique, and only using the GPU capability to do the SW computations one by one. Hence, in this paper, we will propose an efficient SW alignment method, called CUDA-SWfr, for the protein database search by using the intratask parallelization technique based on a CPU-GPU collaborative system. Before doing the SW computations on GPU, a procedure is applied on CPU by using the frequency distance filtration scheme (FDFS) to eliminate the unnecessary alignments. The experimental results indicate that CUDA-SWfr runs 9.6 times and 96 times faster than the CPU-based SW method without and with FDFS, respectively.

Corrigendum: NCNet: Deep Learning Network Models for Predicting Function of Non-Coding DNA

[This corrects the article DOI: 10.3389/fgene.2019.00432.].

Deep convolutional neural network with fusion strategy for skin cancer recognition: model development and validation

We aimed to develop an accurate and efficient skin cancer classification system using deep-learning technology with a relatively small dataset of clinical images. We proposed a novel skin cancer classification method, SkinFLNet, which utilizes model fusion and lifelong learning technologies. The SkinFLNet's deep convolutional neural networks were trained using a dataset of 1215 clinical images of skin tumors diagnosed at Taichung and Taipei Veterans General Hospital between 2015 and 2020. The dataset comprised five categories: benign nevus, seborrheic keratosis, basal cell carcinoma, squamous cell carcinoma, and malignant melanoma. The SkinFLNet's performance was evaluated using 463 clinical images between January and December 2021. SkinFLNet achieved an overall classification accuracy of 85%, precision of 85%, recall of 82%, F-score of 82%, sensitivity of 82%, and specificity of 93%, outperforming other deep convolutional neural network models. We also compared SkinFLNet's performance with that of three board-certified dermatologists, and the average overall performance of SkinFLNet was comparable to, or even better than, the dermatologists. Our study presents an efficient skin cancer classification system utilizing model fusion and lifelong learning technologies that can be trained on a relatively small dataset. This system can potentially improve skin cancer screening accuracy in clinical practice.

Integrating object detection and image segmentation for detecting the tool wear area on stitched image

Flank wear is the most common wear that happens in the end milling process. However, the process of detecting the flank wear is cumbersome. To achieve comprehensively automatic detecting the flank wear area of the spiral end milling cutter, this study proposed a novel flank wear detection method of combining the template matching and deep learning techniques to expand the curved surface images into panorama images, which is more available to detect the flank wear areas without choosing a specific position of cutting tool image. You Only Look Once v4 model was employed to automatically detect the range of cutting tips. Then, popular segmentation models, namely, U-Net, Segnet and Autoencoder were used to extract the areas of the tool flank wear. To evaluate the segmenting performance among these models, U-Net model obtained the best maximum dice coefficient score with 0.93. Moreover, the predicting wear areas of the U-Net model is presented in the trend figure, which can determine the times of the tool change depend on the curve of the tool wear. Overall, the experiments have shown that the proposed methods can effectively extract the tool wear regions of the spiral cutting tool. With the developed system, users can obtain detailed information about the cutting tool before being worn severely to change the cutting tools in advance.

Accelerating Multiple Compound Comparison Using LINGO-Based Load-Balancing Strategies on Multi-GPUs

Compound comparison is an important task for the computational chemistry. By the comparison results, potential inhibitors can be found and then used for the pharmacy experiments. The time complexity of a pairwise compound comparison is O(n (2)), where n is the maximal length of compounds. In general, the length of compounds is tens to hundreds, and the computation time is small. However, more and more compounds have been synthesized and extracted now, even more than tens of millions. Therefore, it still will be time-consuming when comparing with a large amount of compounds (seen as a multiple compound comparison problem, abbreviated to MCC). The intrinsic time complexity of MCC problem is O(k (2) n (2)) with k compounds of maximal length n. In this paper, we propose a GPU-based algorithm for MCC problem, called CUDA-MCC, on single- and multi-GPUs. Four LINGO-based load-balancing strategies are considered in CUDA-MCC in order to accelerate the computation speed among thread blocks on GPUs. CUDA-MCC was implemented by C+OpenMP+CUDA. CUDA-MCC achieved 45 times and 391 times faster than its CPU version on a single NVIDIA Tesla K20m GPU card and a dual-NVIDIA Tesla K20m GPU card, respectively, under the experimental results.

Author Correction: GPU-Accelerated GLRLM Algorithm for Feature Extraction of MRI

An amendment to this paper has been published and can be accessed via a link at the top of the paper.

Diagnostic accuracy of pre-operative breast magnetic resonance imaging (MRI) in predicting axillary lymph node metastasis: variations in intrinsic subtypes, and strategy to improve negative predictive value-an analysis of 2473 invasive breast cancer patients

BACKGROUND

The value and utility of axillary lymph node (ALN) evaluation with MRI in breast cancer were not clear for various intrinsic subtypes. The aim of the current study is to test the potential of combining breast MRI and clinicopathologic factors to identify low-risk groups of ALN metastasis and improve diagnostic performance.

MATERIAL AND METHODS

Patients with primary operable invasive breast cancer with pre-operative breast MRI and post-operative pathologic reports were retrospectively collected from January 2009 to December 2021 in a single institute. The concordance of MRI and pathology of ALN status were determined, and also analyzed in different intrinsic subtypes. A stepwise strategy was designed to improve MRI-negative predictive value (NPV) on ALN metastasis.

RESULTS

2473 patients were enrolled. The diagnostic performance of MRI in detecting metastatic ALN was significantly different between intrinsic subtypes (p = 0.007). Multivariate analysis identified tumor size and histologic type as independent predictive factors of ALN metastases. Patients with HER-2 (MRI tumor size ≤ 2 cm), or TNBC (MRI tumor size ≤ 2 cm) were found to have MRI-ALN-NPV higher than 90%, and these false cases were limited to low axillary tumor burden.

CONCLUSION

The diagnostic performance of MRI to predict ALN metastasis varied according to the intrinsic subtype. Combined pre-operative clinicopathologic factors and intrinsic subtypes may increase ALN MRI NPV, and further identify some groups of patients with low risks of ALN metastasis, high NPV, and low burdens of axillary disease even in false-negative cases.

An Automatic Bleeding-Rank System for Transurethral Resection of the Prostate Surgery Videos Using Machine Learning

Benign prostatic hyperplasia (BPH) is the main cause of lower urinary tract symptoms (LUTS) in aging males. Transurethral resection of the prostate (TURP) surgery is performed by a cystoscope passing through the urethra and scraping off the prostrate piece by piece through a cutting loop. Although TURP is a minimally invasive procedure, bleeding is still the most common complication. Therefore, the evaluation, monitoring, and prevention of interop bleeding during TURP are very important issues. The main idea of this study is to rank bleeding levels during TURP surgery from videos. Generally, to judge bleeding level by human eyes from surgery videos is a difficult task, which requires sufficient experienced urologists. In this study, machine learning-based ranking algorithms are proposed to efficiently evaluate the ranking of blood levels. Based on the visual clarity of the surgical field, the four ranking of blood levels, including score 0: excellent; score 1: acceptable; score 2: slightly bad; and 3: bad, were identified by urologists who have sufficient experience in TURP surgery. The results of extensive experiments show that the revised accuracy can achieve 90, 89, 90, and 91%, respectively. Particularly, the results reveal that the proposed methods were capable of classifying the ranking of bleeding level accurately and efficiently reducing the burden of urologists.