Kohonen’s Algorithm Applied to the Scintigraphic Image for an Aid in the Diagnosis of Prostate Cancer Metastasis

To partition the scintigraphic image, several methods are used, among which is Kohonen’s self-organizing map algorithm. The objective of this study was to perform an ascending hierarchical classification (HAC) on the results of the Kohonen self-organizing map. This makes it possible to carry out the second phase necessary for the elaboration of the classifier by grouping the neurons as well as possible into 3 classes then by reconstituting the scintigraphic image from the 3 classes. This partition proceeds by successive groups, thus merging at each iteration two subsets of neurons using a measure of similarity which is Ward’s method. In this method, the algorithm aggregates the nearest neurons into classes. This allows us to obtain a dendrogram that looks like a tree. And this one needs to be cut. And to have an adequate cut-off level, we have established the variation of the Davies Bouldin index as a function of the number of classes. The minimum value of this index gave the optimal number of classes which corresponded to 3 in the study. These three groups A, B, C have a variable intensity. This intensity can be high, it can be medium or low. The high, medium and low intensities corresponded respectively to metastases for class A, to degenerative or inflammatory phenomena for class B and to normal radiopharmaceutical uptake for class C. To confirm this strong suspicion, we performed reconstructions using a filter. And after this reconstruction, we had images like at the entrance. And for the interpretation of these images, we used a visual metric. This enabled us to note that for the interval [0 - 50[, the image is not contrasted and no lesion could be detected. Over the interval [50 - 200[, we observed the distribution of the radiopharmaceutical over the entire skeletal whole body. On this reconstruction interval, the visual metric shows hypofixation in the bladder and areas suspected of metastases. Over the interval [200 - 250[, we detected hyperfixations linked to degenerative, inflammatory or metastatic lesions. And finally, in the last interval, [250 - 252], we found regions that showed strong uptake (bladder, sternum, etc.). This capture is physiological. Apart from physiological hyperfixation, the other types of hyperfixation were considered metastatic according to the two nuclear scientists who interpreted these images. In total, the HAC allowed us to sub-classify the data into 3 groups which were subsequently reconstructed. And this reconstruction technique highlighted the periarticular metastases belonging to the class [250 - 252]. This allowed us to highlight the oligo-metastases and to carry out in most of these patients a radical prostatectomy.