1.8 METODOLOGÍA DE LA INVESTIGACIÓN
1.8.7. TÉCNICAS E INSTRUMENTO PARA LA RECOLECCIÓN DE DATOS
impact the classification accuracy positively i.e. they increase the accuracies, but certain subbands that have a relatively high discrimination power are seen to decrease the overall classification accuracy. This is due to the fact that a sub- band that is good for differentiating between two subtypes may not be as good for discriminating between all the other meningioma subtypes. Therefore, the computation of discrimination power is susceptible to the bias introduced due to a high distance value between two textures under study and low values for comparison between other texture types. This problem would be compounded as the number of textures for comparison was increased. Hence, the technique is suited to instances where the number of texture classes under study is limited. The results are near perfect for most two class cases using meningioma subtype images.
In future work, we will try to overcome this problem and make the technique more efficient. There are various avenues of further research that may be explored in order to improve the technique and the classification accuracy obtained using it.
6.4
Future Directions
There are various future directions in which the research may be carried out. These are presented in a summary form as follows:
1. Averaging for MAWTT construction: In this thesis we have explored two means of averaging of subband coefficients to construct the MAWTTs. As stipulated in the chapter 3, there could be various other mechanisms to acquire the templates. These could be various other averaging approaches that use different averaging windows. In averaging the size of the window would be an important factor as a small window would be more suited to frequent changes in coefficient values whereas a long window would be more suited for instances where the coefficient values are more consistent. These averaging windows could be different for the various subbands depending upon the frequency content they represent which in turn depends upon the texture samples being analysed. Hence, the search for the optimal averaging
6.4 Future Directions
window could be a subject of further research. Something conclusive can be ascertained with thorough analysis. Although it is important to realize here that the ultimate aim is to obtain stable decompositions which represent subbands that can be used to attain high classification accuracies.
2. Distance functions analysis: In this thesis, we have presented various dif- ferent distance functions and obtained ADWPT representations using each for classifying meningioma subtypes. More distance functions may be used with an emphasis on obtaining a more accurate estimation of the discrimina- tion power of each subband. The important issue to resolve is the accurate estimation of discrimination power for differentiating between multiple tex- ture classes, so that a high distance value for two texture classes does not bias the discrimination power. It may also be found that out of the var- ious short listed distance functions included in the classification analysis in chapter 4 different distance functions may be found better for different texture classification problems.
3. Binary classification and combination: Currently we have tried to compare all the various meningioma subtypes at the same time. A viable alternative would be to compare them two at a time i.e. computation of the various decompositions and classification using the various classifiers. This would raise the issue of ties and generation of multiple ADWPT decompositions for different pairs of meningioma textures. However, more elaborate analysis and experimentation may be carried out to explore all possible modalities of ADWPT and classifier combination in a binary classification mode.
4. Segmentation and structural analysis: Segmentation has been a more pop- ular approach for classification in the domain of histological image analysis with textural analysis not being explored more vigorously. For meningioma subtype classification, segmentation may provide information that could fa- cilitate meningioma subtype classification. Various segmentation schemes may be applied to extract structural information from the histological image samples and these may be combined with our textural analysis technique for better classification results.
6.4 Future Directions
5. Colour information: In our study, we have not explored colour information at all for image analysis. Colour based techniques such as colour features, colour histograms and colour wavelets may be applied to Meningioma sub- type classification.
This thesis presents a novel wavelets based technique that is robust to intra- class variation and achieves efficient feature selection for high accuracy classi- fication results in image classification problems that suffer with low inter-class differences such as Meningioma subtype classification.
Appendix A
List of Publications
H Qureshi, N. Rajpoot, T. Nattkemper, Volkmar Hans, ”A Robust Adaptive Wavelet-based Method for Classification of Meningioma Histology Images”, in
Proceedings MICCAI’2009 Workshop on Optical Tissue Image Analysis in Mi- croscopy, Histology, and Endoscopy (OPTIMHisE), London (UK), 2009
H Qureshi, O. Sertel, R. Wilson, N. Rajpoot, M. Gurcan, ”Adaptive Discriminant Wavelet Packet Transform and Local Binary Patterns for Meningioma Subtype Classification”, in Proceedings Medical Image Computing and Computer Assisted Intervention (MICCAI’08), New York (USA), 2008
H Qureshi, R. Wilson, N. Rajpoot, ”Optimal Wavelet Basis for Wavelet Packets based Meningioma Subtype Classification”, inProceedings Medical Image Under- standing and Analysis (MIUA’08), Dundee (UK), July 2008
H Qureshi, N Rajpoot, R Wilson, T Nattkemper, V Hans, ”Comparative analysis of Discriminant Wavelet Packet Features and Raw Image Features for Classifica- tion of Meningioma Subtypes”, inProceedings Medical Image Understanding and Analysis (MIUA’07), Aberystwyth (UK), July 2007
H Qureshi, NM Rajpoot, K Masood, V Hans, ”Classification of Meningiomas us- ing Discriminant Wavelet Packets and Learning Vector Quantization”, inProceed- ings Medical Image Understanding and Analysis (MIUA’06), Manchester (UK),
July 2006
Khalid Masood, Nasir Rajpoot, Hammad Qureshi and Kashif Rajpoot, ”Hy- perspectral Texture Analysis for Colon Tissue Biopsy Classification”, in Interna- tional Symposium on Health Informatics and Bioinformatics, Turkey 2007
H Qureshi, NM Rajpoot, K Masood, V Hans, ”Classification of Meningiomas us- ing Discriminant Wavelet Packets and Learning Vector Quantization”, in Poster Session of Wave 2006, EPFL, Lausanne, Switzerland, July 2006
H Qureshi, N Rajpoot, T Nattkemper, V Hans, ”Comparison: Meningioma Clas- sification using Wavelet Packets and Normal Texture based Classification”, in1st British Machine Vision Association Student Chapters Meeting, London (UK), March 2007