IMAGE UNDERSTANDING OF MOLAR PREGNANCY BASED ON ANOMALIES DETECTION

Cancer occurs when normal cells grow and multiply without normal control. As the cells multiply, they form an area of abnormal cells, known as a tumour. Many tumours exhibit abnormal chromosomal segregation at cell division. These anomalies play an important role in detecting molar pregnancy cancer.
Molar pregnancy, also known as hydatidiform mole, can be categorised into partial (PHM) and complete (CHM) mole, persistent gestational trophoblastic and choriocarcinoma. Hydatidiform moles are most commonly found in women under the age of 17 or over the age of 35. Hydatidiform moles can be detected by morphological and histopathological examination. Even experienced pathologists cannot easily classify between complete and partial hydatidiform moles. However, the distinction between complete and partial hydatidiform moles is important in order to recommend the appropriate treatment method. Therefore, research into molar pregnancy image analysis and understanding is critical.
The hypothesis of this research project is that an anomaly detection approach to analyse molar pregnancy images can improve image analysis and classification of normal PHM and CHM villi. The primary aim of this research project is to develop a novel method, based on anomaly detection, to identify and classify anomalous villi in molar pregnancy stained images.
The novel method is developed to simulate expert pathologists’ approach in diagnosis of anomalous villi. The knowledge and heuristics elicited from two expert pathologists are combined with the morphological domain knowledge of molar pregnancy, to develop a heuristic multi-neural network architecture designed to classify the villi into their appropriated anomalous types.
This study confirmed that a single feature cannot give enough discriminative power for villi classification. Whereas expert pathologists consider the size and shape before textural features, this thesis demonstrated that the textural feature has a higher discriminative power than size and shape.
The first heuristic-based multi-neural network, which was based on 15 elicited features, achieved an improved average accuracy of 81.2%, compared to the traditional multi-layer perceptron (80.5%); however, the recall of CHM villi class was still low (64.3%). Two further textural features, which were elicited and added to the second heuristic-based multi-neural network, have improved the average accuracy from 81.2% to 86.1% and the recall of CHM villi class from 64.3% to 73.5%. The precision of the multi-neural network
II
has also increased from 82.7% to 89.5% for normal villi class, from 81.3% to 84.7% for PHM villi class and from 80.8% to 86% for CHM villi class.
To support pathologists to visualise the results of the segmentation, a software tool, Hydatidiform Mole Analysis Tool (HYMAT), was developed compiling the morphological and pathological data for each villus analysis.