Background and aim: Determining treatment for prostate cancer involve tissue assessment by grading cancer tissue according to its aggressiveness. Morphological structures of cancer tissue are highly heterogeneous making grading prone to inter observer variability and thus wrongful treatment. To mitigate inter observer variability objective measures are needed. The heterogeneous morphological structures of H\&E stained WSIs provide an opportunity for using CNNs to learn a cancer appearance feature space, that naturally separates cancerous tissue into gleason score. Thus, the aim of this study was to investigate if the gleason score can be determined from the feature space of a supervised CNN without using grade labels. Method: Patches extracted from H\&E stained WSIs were used in a multi output CNN consisting of reconstruction and multi class classification for learning corresponding tissue features. Two model configurations were trained and tested to compare tissue feature separation when grade labels are used. The model's performance was validated using unseen test images with an mean squared error metric for the reconstruction and confusion matrix, precision, recall and F1-score for the multi class classification. For quantifying a gleason score the of the model was extracted, where a principle component analysis was conducted using the features corresponding to >90\% variance. The point-to-point-score algorithm was developed to quantify a gleason score by calculating the label distribution as a function of accumulated euclidean distance of K-means cluster centroids from benign to gleason score tissue features. Results: The mean feature value difference from benign and gleason score 3, 4 and 5 using the test images were (0.005, 0.059), (-0.041, 0.147), (-0.272, 0.187) and (-0.202, 0.153), (-0.290, 0.153), (-0.438, 0.170) with and without grade labels respectively. Using 10 and 25 k-means clusters, the majority of benign, gleason score 3 and 4, and 5 were present between 0 to 0.6, 0.6 to 0.9 0.9 to 1.2 (1.4 -25 clusters). Conclusion: Superior tissue feature separation can be obtained without using grade labels, where the PPS-algorithm can imply a gleason score based on accumulated euclidean distance from benign tissue features.