Lung cancer is the leading cause of cancer death worldwide. One of the reasons is the absence of symptoms at an early stage, which means that it is only discovered at a later stage, where the treatment is more difficult . Furthermore, when making a diagnosis, frequently done by reading computed tomographies (CT's), it is regularly allied with errors. One of the reasons is the variation of the opinion of the doctors regarding the diagnosis of the same nodule [2,3].
The use of CADx, Computer-Aided Diagnosis, systems can be a great help for this problem by assisting doctors in diagnosis with a second opinion. Although its efficiency has already been proven , it often ends up not being used because doctors can not understand the “how and why” of CADx diagnostic results, and ultimately do not trust the system . To increase the radiologists' confidence in the CADx system it is proposed that along with the results of malignancy prediction, there are also results with evidence that explains those malignancy results.
There are some visible features in lung nodules that are correlated with malignancy. Since humans are able to visually identify these characteristics and correlate them with nodule malignancy, one way to present those evidence is to make predictions of those characteristics.
To have these predictions it is proposed to use deep learning approaches. Convolutional neural networks had shown to outperform the state of the art results in medical image analysis . To predict the characteristics and malignancy in CADx system, the architecture HSCNN, a deep hierarchical semantic convolutional neural network, proposed by Shen et al. , will be used.
The Lung Image Database Consortium image collection (LIDC-IDRI) public dataset is frequently used as input for lung cancer CADx systems. The LIDC-IDRI consists of thoracic CT scans, presenting a lot of data’s quantity and variability. In most of the nodules, this dataset has doctor’s evaluations for 9 different characteristics. A recurrent problem in those evaluations is the subjectivity of the doctors' interpretation in what each characteristic is. In some characteristics, it can result in a great divergence in evaluations regarding the same nodule, which makes the inclusion of those evaluations as an input in CADx systems not useful as it could be.
To reduce this subjectivity, it is proposed the creation of a metric that makes the characteristics classification more objective. For this, it is planned bibliographic and LIDC-IDRI dataset reviews. With that, taking into account this new metric, validated after by doctors from Hospital de São João, will be made a reclassification in LIDC-IDRI dataset. This way it could be possible to use as input all the relevant characteristics.
The principal objective of this dissertation is to develop a lung nodule CADx system methodology which promotes the confidence of specialists in its use. This will be made classifying lung nodules according to relevant characteristics to diagnosis and malignancy. The reclassified LIDC-IDRI dataset will be used as an input for CADx system and the architecture used for predicting the characteristics and malignancy results will be the HSCNN. To measure the classification evaluation will be used sensitivity, sensibility, and area under the Receiver Operating Characteristic (ROC), curve.
The proposed solution may be used for improving a CADx system, LNDetector, currently in development by the Center for Biomedical Engineering Research (C-BER) group from INESC-TEC in which this work will be developed.
 - S. Sone M. Hasegawa and S. Takashima. Growth rate of small lung cancels detected on mass ct screening. Tire British Journal of Radiology, pages 1252–1259
 - D. J. Bell S. E. Marley P. Guo H. Mann M. L. Scott L. H. Schwartz D. C. Ghiorghiu B. Zhao, Y. Tan. Exploring intra-and inter-reader variability in uni-dimensional, bi-dimensional, and volumetric measurements of solid tumors on ct scans reconstructed at different slice intervals. European journal of radiology 82, page 959–968, 2013
 - H.T Winer-Muram. The solitary pulmonary nodule 1. Radiology, 239, pages 39–49, 2006.
 - R. Yan J. Lee L. C. Chu C. T. Lin A. Hussien J. Rathmell B. Thomas C. Chen et al. P. Huang, S. Park. Added value of computer-aided ct image features for early lung cancer diagnosis with small pulmonary nodules: A matched case-control study. Radiology 286, page 286–295, 2017
 - W Jorritsma, Fokie Cnossen, and Peter Van Ooijen. Improving the radiologist-cad interaction: Designing for appropriate trust. Clinical Radiology, 70, 10 2014.
 - Tom Brosch, Youngjin Yoo, David Li, Anthony Traboulsee, and Roger Tam. Modeling the variability in brain morphology and lesion distribution in multiple sclerosis by deep learning. Volume 17, 09 2014. - Simon Aberle Deni A. T. Bui Alex Hsu Willliam Shen, Shiwen X. Han. An interpretable deep hierarchical semantic convolutional neural network for lung nodule malignancy classification. june 2018
Nº de páginas: