EE Seminar: Enhancing Transfer Learning for Pulmonary Nodule Detection using Preprocessing Techniques
Speaker: Max Fomin
M.Sc. student under the supervision of Prof. Hayit Greenspan
Wednesday, December 19th, 2018 at 15:30
Room 011, Kitot Bldg., Faculty of Engineering
Enhancing Transfer Learning for Pulmonary Nodule Detection using Preprocessing Techniques
One of the biggest challenges today in implementing deep learning systems in the medical world is obtaining sufficiently large datasets that enable training neural networks without overfitting. One notices that the state-of-the-art systems in everyday object detection tasks are trained on huge datasets, in order to achieve their accuracy. In contrast, the medical world, which inherently requires the highest accuracy because of the high cost of errors, suffers from smaller datasets by far, dictating the creation of much simpler neural networks that thus achieve worse results.
Exisiting CAD (Computer Aided Diagnosis) systems for pulmonary nodule detection have two main stages. The first stage is the nodule candidate generation, which aims to produce as many high quality candidates as possible using 2D slices for fast runtimes. The second stage is a false-positive filtering stage, which aims to pass only the candidates that are true nodules. This stage, which operates only on the candidates, and not on the entire slice, can thus work in higher dimensions, i.e. 2/2.5/3 dimensional object classification.
Our work is the introduction of the MiMax Technique, a unique pre-processing method for medical images that improves the transfer learning process from public datasets of everyday images. In order to give a theoretical introduction to this method, we also present the SPCLAHE technique.
The SPCLAHE method is a proven pre-processing method for boosting the analytics of medical images (specifically, malignancy detection in mammographies) using CNNs. A significant advantage of SPCLAHE is that its product is a color image. Eventually, we introduced the MiMax Technique, our novel contribution in this work, presenting the best results. The MiMax Technique essentially fuses together the CLAHE algorithm from the SPCLAHE method with the jet colormap, including the best of all worlds. Running the MiMax Technique resulted in up to 14% improvement in object detection performance, a very impressive result given that we have not changed any parameters in the neural network itself or in its training procedure.
Our work is based on 2D object detection and therefore the aim of its product is to improve the candidate generation stage in these systems. The idea is plugging this method into any existing CAD system, in order to boost its performance, without the need of changing the system itself.
