350 rub
Journal Information-measuring and Control Systems №10 for 2013 г.
Article in number:
Neural network detection of tumors on mammograms using the MSER algorithm and texture features
Authors:
A.V. Dolgopolov - Senior programmer-mathematician, Pawlin Technologies Ltd. E-mail: avido@pawlin.ru
P.A. Kazantsev - Head of the development Department, Pawlin Technologies Ltd. E-mail: pak@pawlin.ru
Abstract:
The problem of automatic tumor detection on mammograms concludes in a fact that tumor lacks formalizable structure. Various detection algorithms use information about local structural elements of an object, comparison of brightness characteristics of object-s zones with each other, etc. For tumor detection on mammograms such approaches, generally, will fail due to the reason stated above, especially in a case of small tumors typical for early stages of disease, which are of the most importance for timely tumor diagnostics, when chances of successful treatment are highest. Tumors on mammograms appear either as the brightest or the darkest regions of the image. Regions with similar characteristics of brightness and form are scattered across the whole mammogram. Let us denote these regions as regions of interest (ROI). Additional information can be extracted from the ROI-s surroundings and then used for classification. In a proposed algorithm, ROIs are extracted using MSER (Maximally stable extremal regions) algorithm. Currently, MSER appear to be a popular method for detection of image regions that significantly differ from their surroundings by brightness characteristics. The difference of brightness characteristics in ROI and its surroundings is not enough for acceptable precision of classification. In a proposed algorithm additional information is extracted from texture of ROI and its surroundings. For this, we use local binary patterns operator (LBP). LBP-transform is applied to the whole image, and then, for each ROI, we build several LBP-histograms. To do this, we divide each ROI into three sub-regions, which we denote as central, inner and outer. LBP-histogram is represented by a vector of values, that, potentially, can be of a high dimensionality. In order to counter problems caused by a "curse of dimensionality", that especially manifests itself in cases of small training set, for each sub-region we select only few histogram components that bear the most information. To do so, we perform a cycle of preliminary feed-forward neural network training. At each iteration of a cycle we add a histogram component to the feature vector, compute current error and compare it with an error on a previous iteration. Cycle stops when error drop is not sufficient enough (regulated parameter). Resulted feature vector that corresponds to iteration with least error is used for further extensive training. Feature vector found was then used for experiments which aim was to find the best neural network architecture. The most optimal network configuration found was a three-layered neural network - 8x80x1. Output value ranges from 0 to 1 and corresponds to probability of current ROI being a tumor. Training, testing and validation of neural network classifier were performed using database of mammograms marked up by a radiologist. Quality testing was carried out using ROC-curve analysis. Recall and FPR was measured mammogram-sample-wise (considering the fact that one mammogram in a set can contain only one tumor), not a ROI-wise. Area under ROC-curve was computed as well. Experiments have shown the following quality characteristics: recall - 1.0, FPR - 0.08, AUC - 0.89.
Pages: 66-69
References

  1. Matas J., Chum O., Urban M., Pajdla T. Robust wide baseline stereo from maximally stable extremal regions». Proc. of British Machine Vision Conference. 2002. P. 384-396.
  2. Mikolajczyk K., Tuytelaars T., Schmid C., Zisserman A., Kadir T. and Van Gool L. A Comparison of Affine Region Detectors // International Journal of Computer Vision. 2005. V. 65. № 1-2. November. P. 43-72.
  3. Dalal N., Triggs B. Object Detection using Histograms of Oriented Gradients. Pascal VOC 2006. Workshop. ECCV 2006. Graz, Austria.
  4. Unay D., Ekin A., Jasinschi R.S. Local Structure-Based Region-of-Interest Retrieval in Brain MR Images // Information Technology in Biomedicine. IEEE Transactions on. 2010. V.14. № 4. P. 897, 903.
  5. Unay D., Ekin A., Cetin M., Jasinschi R., Ercil A. Robustness of Local Binary Patterns in Brain MR Image Analysis // Engineering in Medicine and Biology Society. 2007. EMBS 2007. 29th Annual International Conference of the IEEE. 2007. P.2098, 2101.
  6. Zhen Lei, Chan A.K. An artificial intelligent algorithm for tumor detection in screening mammogram // Medical Imaging. IEEE Transactions on. 2001. V.20. № 7. P. 559,  567.
  7. Cheng H.D., Shi X.J., Min R., Hu L.M., Cai X.P., Du H.N. Approaches for automated detection and classification of masses in mammograms. Pattern Recognition. 2006. April. V. 39. Issue 4. P. 646-668.