Presenter

Mr Alexander Soñora Mengana - VUB-ETRO [Email]

Abstract

Lung cancer is the first cause of cancer related death worldwide Early detection can have substantial impact on treatment outcome Computer aided detection (systems can play an important role in improving the detection rate and reducing the clinical workload, in particular considering the lung cancer screening protocols that are currently being set up.
Starting from an existing system for computer aided detection of lung cancer, several aspects of the processing pipeline were investigated, with the aim to improve the accuracy and robustness of the process The system employed a two stage approach, comprising of a candidate detector, and a false positive reduction step based on hand crafted features Initially the design of the system was changed to a modular architecture to facilitate introducing alterations at different stages, and evaluate their impact The system's efficiency and usability was improved and individual components were tuned.
Next, a thorough characterisation of its performance by participating to the LUNA 16 Challenge The participation implied training and testing the system on large clinical dataset It also enabled the objective comparison to other proposed CAD approaches using a common evaluation methodology The system as a whole, was shown to perform well, achieving comparable results to other full system submissions at the time of the challenge Closer analysis, revealed this was mainly due to a sensitive nodule candidate detector, whereas other approaches were found to have better false positive reduction.
Subsequently, several aspects of the pipeline were investigated, to improve on this baseline results An improved lung segmentation procedure was added to the preprocessing stage The method reduces the amount of failed lung segmentations due to artefacts or even tracheotomy by performing an error detection and correction procedure, making the candidate detection process more robust.
Candidate detectors often mark a large number of non nodule structures compared to the of actual nodules This imbalance in the data during training, may hinder the performance of the classifier Data balancing methods, comprising both undersampling and oversampling approaches in feature space, were therefore investigated in detail Surprisingly, undersampling the majority class, as performed in the original system, was found to perform worse compared to no balancing Balancing by oversampling the minority class allowed to improve the result further
Over the course of my PhD, deep learning methods emerged for medical image analysis, and rapidly outperformed alternative approaches in the CAD domain I therefore investigated how to increase the accuracy of the false positive reduction by training a convolutional neural network using the candidates provided by my detector, and obtained a substantial increase in accuracy
Interestingly, combining the learnt features with the hand crafted features, improved the results even further

https://meet.jit.si/defensapublicadetelecomunicaionesyelectronica