- Title
- Machine learning-based lung nodule detection on chest x-ray radiographs
- Creator
- Li, Xuechen
- Relation
- University of Newcastle Research Higher Degree Thesis
- Resource Type
- thesis
- Date
- 2016
- Description
- Research Doctorate - Doctor of Philosophy (PhD)
- Description
- Lung cancer is one of the most deadly diseases. Worldwide it has the highest rate of incidence and death of any cancer. Early diagnosis of lung cancer is the key to providing the best possible clinical outcomes for patients. As an initial diagnostic tool for a variety of clinical conditions, chest x-ray (CXR) radiography is the most commonly used radiological examination by far, making up at least a third of all examinations in a typical radiology department. Many diseases such as lung cancer can be diagnosed at an early stage by regular health screening, and since lung cancer can appear as a solitary nodule in chest x-ray images, the detection of lung nodules in chest x-ray images could have a significant impact on the early detection of lung cancer. In this thesis, machine learning-based methods were used for lung nodule detection in CXR images. Three studies were conducted, including lung field segmentation, rib recognition and suppression, and lung nodule detection. The inclusion of these three studies is based on the following diagnostic considerations. First, the lung field is the area of interest in chest radiographs for lung disease diagnosis. Since the size, shape and texture of a lung field are key parameters for lung disease diagnosis, it is necessary to segment the lung field from x-ray chest images. Second, in chest radiographs, bones and organs are overlapping in the image. In contrast with computed tomography (CT) or magnetic resonance imaging (MRI) images, the image data of X-ray images are two dimensional, having a high level of noise due of the overlapping. In lung nodule detection, the main noise is that of the ribs that overlap the lung field and tumours. Studies have shown that 30% of nodules present in chest radiographs were missed by radiologists that were visible in retrospect; and 82% to 95% of the missed nodules were partly obscured by overlying bones such as ribs and clavicles. It is therefore necessary to locate the ribs and suppress them to reduce their obscuring effect. When the ribs on CXR images are removed, the nodules become more visible. Third, CXR images are traditionally diagnosed by radiologists and doctors. However, in some developing countries or some underdeveloped areas, hospitals lack experienced radiologists. A lung nodule detection system can help them to discover inconspicuous lung nodules. On the other hand, in some populous countries, when doing general health screening, there will be numerous CXR images that need to be diagnosed in a short time. A lung nodule detection scheme can be used as a second opinion for assisting radiologists’ decision-making to avoid overlooking more subtle nodules. Based on these three studies, an automatic lung nodule detection system has been developed. Firstly, an advanced statistic shape and appearance model method for lung field segmentation was developed. The traditional multi-resolution methods for statistic shape models were optimised and a new combination of features for the appearance model was proposed. Secondly, a novel rib recognition method and an advanced rib suppression method were proposed based on principal component analysis (PCA). For rib recognition, firstly a graph-based method was employed for rib recognition on CXR images; then machine learning-based methods were utilised to replace the traditional rule-based methods to refine the rib recognition results. For rib suppression, it was necessary to solve the problem that the PCA-based method cannot suppress rib edges. By detecting the edge of rib model, the edges of ribs can be suppressed by using neighbouring pixels. Thirdly, a wavelet transformation and a convergence index filter were employed to extract the texture features of CXR images, with AdaBoost (a multiple weak learners boosting method) being used to identify lung nodule candidates. A novel parameter to evaluate the stationary degree of lung nodule candidates was proposed, and by using this parameter, the lung nodule detection performance was greatly improved. The system could achieve similar performance to that of a doctor of average experience in lung nodule detection. It can therefore be used to help doctors to locate inconspicuous lung nodules and reduce the rate of false negatives. The methods that were proposed in all three studies achieved better performance than previous studies. For the statistical shape and appearance model-based lung field segmentation method, the average overlap of segmentation results and ground truth was above 93%. It outperformed other active shape model-based methods reported so far. The rib recognition using a support vector machine and decision tree could recognize 90% of visible ribs the in lung field. The principal component analysis-based rib suppression method is shown to remove most rib textures and achieve similar suppression performance as that of dual-energy subtraction technology. The machine learning-based lung nodule detection method is shown to detect over 94% of the lung nodules that were in the lung fields in the database developed by Japanese Society of Radiological Technology (JSRT) where the false positives per image was 5. These experiments on lung nodule detection have shown that the proposed system has the potential to be used in clinical practice.
- Subject
- CXR; machine learning; lung nodule detection; lung field segmentation; rib suppression
- Identifier
- http://hdl.handle.net/1959.13/1313753
- Identifier
- uon:22638
- Rights
- Copyright 2016 Xuechen Li
- Language
- eng
- Full Text
- Hits: 867
- Visitors: 2271
- Downloads: 1334
| Thumbnail | File | Description | Size | Format | |||
|---|---|---|---|---|---|---|---|
| View Details Download | ATTACHMENT01 | Abstract | 218 KB | Adobe Acrobat PDF | View Details Download | ||
| View Details Download | ATTACHMENT02 | Thesis | 2 MB | Adobe Acrobat PDF | View Details Download |