ABSTRACT Image muscle segmentation is useful to quantitatively assess musculoskeletal diseases by extracting biomarkers such as shape, texture and water diffusivity metrics. Although volumetric manual segmentation is time consuming and a bottleneck in practice, fully automatic approaches are still in progress to reach an acceptable accuracy. In this paper, we provide a robust semi-automated tool to segment two musculoskeletal systems, i.e. thigh and shoulder in MRI and CT modalities, respectively. The tool only needs a few manually labelled cross-sections to build a directed graph-structure of corresponding points between the successive spaced slices. The boundaries of each muscle are obtained by performing a spline interpolation based on the directed graph-structure. Each muscle label and its corresponding 3D mesh are deduced using post-processing techniques. We evaluated the tool on 26 MRI thighs and 16 CT shoulders. Three metrics along with inter-muscle overlapping were employed to evaluate the tool by comparison to an expert manual segmentation and a publicly available tools (ITK-SNAP, 3D Slicer). The results showed a mean Dice 0.988 ± 0.003 , and Hausdorff Distance 4.86 ± 1.67 mm in comparison to the manual reference for thigh muscle segmentation, and a mean Dice 0.961 ± 0.005 and Hausdorff Distance 2.42 ± 0.79 mm for shoulder muscle segmentation, outperformed the other methods. The tool is proposed as slicer module available at https://github.com/latimagine/SlicerSpline.

ABSTRACT The aims of this study were 1) to propose a simple two-dimensional (2D) video-based method to measure segment lengths in the clean and jerk during weightlifting competitions, and 2) to check concurrent validity and test-retest reliability compared to 3D video-based measurements. In the first part of the study, key body positions for the assessment of segment lengths via 2D measurement method were determined. In the second part, concurrent validity was verified, comparing the 2D segment length measurements to 3D measurements. In the third part, the test-retest reliability of the 2D approach was investigated. In the clean and in the jerk, the lift-off/start position and the position of maximum barbell velocity displayed a minimised offset between ‘real’ 3D segment lengths and 2D projections of the segment lengths. For the concurrent validity assessment, the overall random measurement error was 4.3% with concordance correlation coefficients ranging from 0.64 to 0.88. The standard error of measurement in the test-retest procedure was 2.4% on average with intraclass correlation coefficients ranging from 0.85 to 0.96. The proposed 2D video-based approach to measure body segment lengths during weightlifting exercises is proven to be a valid and reliable procedure.

ABSTRACT The COVID-19 pandemic underscores the vital need for accurate lung infection diagnosis to guide effective medical interventions. In response, this research introduces a novel deep multi-task model that seamlessly integrates segmentation and classification tasks for the detection of COVID-19 in CT scan images. This innovative model leverages a shared encoder for feature extraction, a dedicated decoder for segmentation, and a multi-layer perceptron for classification. The primary objective of this model is to address the challenge of task imbalance introduced by the application of image processing algorithms in the multi-task models. Our study involves a two-stage evaluation. Initially, we apply the proposed multi-task model with image processing algorithms to highlight task imbalance. Subsequently, we balance tasks by combining binary image processing algorithms. Evaluation on four datasets shows impressive results with a Dice coefficient of 88.91 ± 0.01 for segmentation and 0.97 classification accuracy. In summary, this model advances medical image analysis for enhanced diagnostic precision in healthcare.

ABSTRACT The liver is the largest solid organ in the body that can be anatomically divided into segments. We present in this work a web-based subject-specific human liver atlas based on the Couinaud segments simulated from portal venous (PV) perfusion zones, hepatic arterial (HA) and hepatic venous (HV) trees, as well as biliary drainage. The purpose of the atlas is to provide the modelling community with freely accessible 3D hepatic structures for in silico simulations, which are of tremendous value in yielding novel insights in hepatic circulation, drug transport and clearance.

ABSTRACT Indices derived from photoplethysmography (PPG) have shown promising results as non-invasive digital biomarkers for the detection of diabetes mellitus (DM). Considering the mutual endothelial insult leading to similar undesirable peripheral hemodynamic perturbations, hypertension (HT) may blunt this classification performance. Second derivative PPG (SD-PPG) indices were derived from the second derivative of the PPG signal. The variables of interest were the previously described peaks of the initial positive (a), early negative (b), re-increasing (c), late re-decreasing (d), diastolic positive (e) and negative (f) waves and the ratios between them. Patients were classified according to their type 2 DM and hypertension phenotypes. SD-PPG indices were compared between diseased subgroups, healthy controls and also dichotomous classification performance was evaluated. Two SDPPG indices, b/a ratio and the vascular ageing index (VAI = (b-c-d-e)/a) responded to isolated DM type 2 (n = 29) amongst healthy subjects (n = 106) (area under the curve (AUC) = 0.629 p = 0.034 and 0.631 p = 0.031 20 respectively). However, the classification performance became insignificant with the inclusion of HT patients (n=30). (p = 0.839 vs. p = 0.656). These results suggest that the coexistence of HT and DM may hinder the use of SD-PPG for noninvasive DM detection.

ABSTRACT A virus known as the Coronavirus 2019 (COVID-19) can have serious effects on a variety of body parts. Preventing this spread of the virus requires early detection and proper treatment. Many methods are followed for early detection of this disease based on symptoms and classification for proper treatment. In this paper, classification of COVID-19 using Computed Tomography (CT) scan images with hybrid optimisation-enabled deep learning is researched. Here, a hybrid optimisation algorithm called War Snake Optimisation Algorithm (WSOA) is proposed to train Deep Q-Network (Deep Q-Net). Moreover, the proposed WSOA is formed by integrating War strategy optimisation (WSO) and Snake Optimiser (SO). In this paper, pre-processing is carried out using a median filter, followed by lung lobe segmentation that is done by Generative Adversarial Network (GAN). Furthermore, image augmentation is done by translation, rotation, flipping and scaling. Then, feature extraction is followed after image augmentation, where needed features get extracted to end up with the classification of COVID-19. This classification brings out the patient’s health condition of normal and abnormal situations. Moreover, this research work is analysed with three performance metrics, such as accuracy, sensitivity and specificity with values of 93.7%, 94.8% and 89.5%.

ABSTRACT Breast cancer is the most common cancer-related death in women, accounting for 16% of all cancer-related fatalities globally. Breast cancer is fatal in just half of all cases. Radiologists may misread worrisome lesions due to imaging quality concerns and diverse breast densities, which raises the false-(positive and negative) ratio, as the primary explanation for the problem. Early intervention is critical in building a current prognosis process that can successfully limit disease consequences and increase recovery. Patients are being referred back for biopsies to dispel suspicions when the inconsistent feature-extraction approach is used for manual screening of breast abnormalities in traditional schemes. With the Multi-kernel Support Vector Machine (MKSVM), we develop a new modality for the prediction of breast cancer and train its properties using the supervised machine learning approaches. Furthermore, the IDA-MKSVM technique achieved maximum average accuracy is 95.75%, average sensitivity is 94.29%, average specificity is 95.16% and average F-score is 95.39% for different training datasets. For selecting ideal features, the Improved Dragonfly Algorithm (IDA) is also used. A 10-fold cross validation procedure is used in the system under consideration to ensure accuracy. The UCI machine learning repository holds the breast cancer diagnosis data referred to as the Wisconsin breast cancer diagnosis data set.

ABSTRACT We present a semi-automatic reconstruction approach of middle and inner ear structures using generic 3D deformable surface models from Cone Beam CT (CBCT) examination. First, the user must position a set of control points in the CBCT volume for each of the 4 structures of the inner and middle ear. These points are used to position the deformable surface models and to customize them so that they are as close to the boundaries as possible. Finally, each mesh is refined iteratively segmenting the limits of the structure while taking into account neighbouring structures as boundary constraints. Our method is tested on left and right ears of 20 scans of patients analysed retrospectively. The results show the efficiency and reliability of this approach with an average Dice Similarity Coefficient of 91.8% for the inner ear model and 89.9% for the ossicular chain and a total reconstruction time of 5 minutes. The implementation of our method in a clinical setting could provide clinicians with distinct and accurate 3D models of the ear structures without requiring a tedious manual segmentation step, in order to give them a better understanding of the auditory system in vivo and help them in diagnosis and follow-up.

ABSTRACT First-pass Effect(FPE) for Endovascular Therapy(EVT) is associated with good clinical outcome (mRS2) of patients with Acute Ischaemic Stroke (AIS). A rapid and accurate prediction of FPE before EVT can help neurointerventionists plan the procedure and avoid delays in restoration of cerebral blood flow. However, there are rare studies focused on FPE prediction on arterial vessel anatomy immediately. The intractable difficulty lies in extracting discriminative features to represent a wide variety of vessels with irregular vessel shapes. In this paper, we propose a Symmetric Positive Definite(SPD) matrix-based feature representation extracted from the centerline of arterial vessels, encoding the global discriminative information over the artery for predicting FPE. Subsequently, a collaborative network of multi-scale Convolutional Neural Network(CNN) and Multiple Layer Perception (MLP) with specific attentions is developed. Specifically, the CNN is used for capturing the features among the multi-scale local neighbours of the curve. MLPs are used for capturing more prominent global discriminative features at different scales. The attention mechanism is used to better filter and extract the useful information for feature fusion. Quantitative experimental results demonstrate that our proposed method is able to predict FPE accurately, outperforming the manually defined features and traditional machine learning-based methods in this task, regarding the metrics of AUC, precision, sensitivity, specificity and accuracy.

ABSTRACT The early detection and treatment of DR have helped the ophthalmologist to treat the affected patients and to reduce vision loss. Computer-aided screening for automatic DR detection in the medical system has consistent detection of lesions in retinal fundus image. To overcome these challenges and to offer timely treatment, this paper aims to develop a novel deep learning-based DR detection. Here, the integration of ‘Optimized Iterative Thresholding (O-IT)’ is adopted for the accurate segmentation of blood vessels. The first novelty of this work is that the thresholding approach is improved by tuning the parameters in the proposed model by developing a hybrid meta-heuristic Shark Smell-Jaya Optimisation (SS-JO) algorithm to enhance the performance of both blood vessel segmentation and classification. CNN is replaced by a deep learning framework termed as optimised ‘Long Short-Term Memory (LSTM)’. The second novelty of this work is that the optimised LSTM is designed in the proposed model by optimising the parameters in LSTM using the implemented SS-JO to reduce the complexity of the network. The accuracy analysis of the implemented SS-JO-CN-LSTM is secured 44%, 29%, 19%, 6%, 4% and 15% improved than SVM, NN, CNN, LSTM, CN-LSTM and FR-CSA-NN+CNN.