In order to better understand how false negatives occur within object detection tasks, current research has identified five ‘false negative mechanisms’. These mechanisms describe how a component within the object detector architecture failed. In particular, the ‘missed ground truth’ mechanism is most consequential due to the risks associated with its occurrence, however it remains largely unexplored within temporal data. I aim to identify how ‘missed ground truth’ errors occur, with a focus towards the variation in pixel size and occlusion of target objects over time. I propose a method that involves training a faster region-based convolutional neural network model to produce predictions of vehicles and pedestrians within an autonomous driving benchmark dataset. Furthermore, I introduced an algorithm to detect and quantify the occurrence of the false negative mechanism. The bounding box surrounding the target object served as the basis for determining the pixel size, whilst the occlusion of the target object was measured by the percentage of which it was covered. To identify trends, plots were used to visualise the collected data. The results indicate that the likelihood of detecting a target object decreases with distance or visual obstruction. A higher false negative rate exists when the distance to the target object increases, however this situation poses minimal risk as there is still opportunity for the object to be detected. Despite the lower false negative rate associated with occluded objects, this situation poses greater risk as the occluded object may not be detected at all, resulting in severe consequences. The approach taken to identify how false negatives occur within temporal data provides a basis to conduct further research into the reliability of robotic vision, as it becomes increasingly implemented in day to day life.
Pelvic Organ Prolapse (POP) is a condition typically resulting from the process of vaginal birth. Diagnosed women often exhibit evidence of injury, via MRI, to their pelvic floor muscles. This induced weakness can result in one or more of the pelvic organs: vagina, cervix, uterus, bladder, urethra and rectum, to drop from their normal positioning, causing abnormal discomfort and pain. Patients have until recently been treated via implantation of polypropylene pelvic meshes to permanently reinforce their pelvic floor muscles. As polypropylene is a nonbiodegradable polymer, this aspect has caused many ongoing health complications such as: mesh erosion into vaginal tissue, chronic bacterial infection, incontinence, pain and severe discomfort. As a result, the United State Food and Drug Administration, and Australian Therapeutic Goods Administration have issued updated regulations, arranging for numerous mesh product recalls and cancellations. Simultaneously, the quantity of pelvic mesh implant removal surgeries has increased. This has provided an opportunity to determine the relationship between inflammation levels of continence and prolapse mesh implants based on the quantity of identified macrophage inflammation markers. This study investigated inflammatory responses to various pelvic floor mesh products. These samples were explanted from patients at both the Royal Brisbane and Women’s Hospital and Wesley hospitals for histological analysis. Such processes included sample fixation, paraffin embedment, microtome sectioning and immunohistochemistry staining of the tissue. Quantification of inflamed interest regions, indicated by an intense brown colouring, were accomplished through image processing via deconvolution of the sample stain colours with a custom MATLAB code. This has provided the basis for quantification of three macrophage markers: General (CD68), M1 (iNOS) & M2 (CD206). The results reveal the relationship between different mesh implants and macrophage expression, which is essential to understanding the primary cause for POP patient’s on-going pain and discomfort post-surgery. Moving forward, this work will support the understanding and improvement of treatment for women suffering from Pelvic Organ Prolapse.
In the health sector, the Gastrointestinal (GI) tract endoscopy procedure plays a major role in identifying any underlying abnormalities within the GI tract of a human. There are multiple GI tract diseases that can be sometimes life-threatening, such as precancerous lesions and other intestinal cancers. In the usual process, this is performed by a medical expert manually which can be prone to human errors and also can entirely depend on the level of experience of medical experts. Therefore, we investigated an AI-based model to automate the abnormality detection process in the endoscopy data. Deep learning, specifically convolution neural networks (CNNs) which are designed to perform automatic feature learning without any prior feature engineering, has recently been reported to be highly beneficial and also made advances in GI endoscopy image analysis. Previous research has developed models that focus only on improving performance (such as accuracy), such that, the majority of introduced models contain complex deep network architectures with a large number of parameters that requires longer training times. However, there is a lack of focus on developing lightweight models which can run on low-resource machines with low-quality datasets.
In this study, we investigate the possibility to develop an efficient GI disease classification method that is both lightweight and achieves considerable performance compared to the existing heavy-weight models. We evaluate our model using the recent Kvasir dataset (i.e. KVASIR-V2) of endoscopic images of gastrointestinal diseases. The dataset comprises 4,000 coloured images labelled and verified by medical endoscopists, covering multiple pathological findings such as esophagitis, polyps, ulcerative colitis, and a few anatomical landmarks: Z-line, pylorus, and cecum. This approach is motivated by the concept of Knowledge Distillation, in which the small network (student model) mimics the larger cumbersome network (teacher model) in order to achieve competitive performance. We designed a lightweight network as our student model and adapted the Res-Net50 as our teacher model, which is a pre-trained CNN (i.e. ResNet50) trained on a large image classification database called ImageNet. However, as the ImageNet database does not contain any medical images we performed transfer learning by fine-tuning the ResNet50 with a sample of medical images from the KVASIR dataset, before utilising it as the teacher network. Then during the training of the lightweight model, we have incorporated the teacher network to provide guidance to the student model, in order to achieve better classification performance.
Except where otherwise noted, content on this site is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Australia License.
QUT acknowledges the Traditional Owners of the lands where QUT now stands.