A visual model approach to extract regions of interest in microscopical images of basal cell carcinoma

The model was tested with a total of 115 histological microscopical fields of view of different types of basal cell carcinoma, sampled from 25 randomly chosen patients. Each biopsy was formalin-fixed and stained with Hematoxylin-Eosin dyes. Microscopical fields were digitized with a Nikon eclipse E600 system, through a coupled Nikon DXM1200 camera, and stored in JPEG format at a 1280 × 1024 resolution using microscope magnifications of ×4, ×10, ×20 and ×40. An expert pathologist, with at least five years of experience, selected the digitized fields of view and manually segmented relevant regions. We use 20 images to extract textons of size 32 × 32 pixels from RoIs and background for the object recognition task, and 95 images to test the entire algorithm.

During a standard exploration of a histopathological sample, the pathologist integrates two types of information sources, namely, 1) the visual field content itself (“bottom-up” source), and 2) the knowledge involved in the specific task of diagnosis (“top-down” source). The HV S fuses together these sources using at least four different brain areas, namely, 1) the V1 cortex which assigns a local relevance to the visual input, 2) the V2 cortex which is responsible of gathering together these relevancies as simple shapes, 3) the V4 cortex which actively regulates the V2 excitation, and finally, 4) the inferotemporal gyrus which integrates the function of the previous areas by recognizing complex shapes and their purposes in the scene and probably by retrieving a slight image representation composed of a few objects actually recognized, the relations between them and the background, and basic information about the background texture [6]. The approach proposed herein models the function of the HV S in four steps, as shown in Figure 1. Firstly, it assigns a local relevance by integrating information from basic features as orientation, color and intensity at multiple scales, as previously described by Itti et al. [7] (V1 cortex function). Secondly, the model segments the image by taking into account the proximity and similarity between pixels and mixing up the conspicuity maps into the resulting regions (V2 cortex function), but also adding a map of the intrinsic structural disorder which models the specific task knowledge that regulates the attention over each structure (V4 cortex function) [5]. Thirdly, this intermediate map is used to feed a module that efficiently looks for basal cell carcinoma by comparing the pattern composition with a data base of carcinoma regions, starting with the highest attentional regions (inferotemporal gyrus function). Once the algorithm sets up the first carcinoma region within the image, the other carcninoma regions are defined as the most similar regions using an Euclidean metrics of the different basic features: color, entropy and image intensity.

Then, this intermediate map is used to feed a module that efficiently looks for basal cell carcinoma by comparing the pattern composition with a data base of carcinoma regions, starting with the highest attentional region (inferotemporal gyrus function). Once the algorithm sets up the first carcinoma region within the image, the other carcninoma regions are defined as the most similar regions using an Euclidean metrics of the different basic features: color, entropy and image intensity.

During a searching task of an undetermined number of targets, the pathologist determine whether she/he is familiar with the scene, using simple features as color and texture information, possibly recognizing some relevant objects and their relations [6, 8]. If this is the case, she/he efficiently looks for the targets, following the saliency of the scene and then using the target contextual relationships [8, 9]. Likewise, our proposal follows the most saliency levels of the map obtained in the local recognition step [5], that is say the third stage, when the system has determined a target. We extract then some textons from RoIs previously marked (object recognition) and some from the background (basic background features extraction) to discard irrelevant structures. A classical kNN algorithm was used, using the low level features captured before as input [10]. In case of finding a target, its relations with the other structures in the image are determined, comparing how different its internal features are. Formally, the distance between features (df ) within the target structure (st ) and the i-structure (si ) is defined as the Euclidean mean distance of the intensity features (I), orientation (O), color (C) and entropy(H) inside each structure, formally, 0 identifying similar structures as part of the searched targets using a simple threshold. Note that this algorithm addresses the intensive computations to a few structures, until it reaches the first target.