Morphometric information to reduce the semantic gap in the characterization of microscopic images of thyroid nodules

• Theoretical–practical study using Computer Aided Diagnosis and Morphometry concepts.
• The method aggregates the content of nuclei in microscopic images.
• The images are characterized by morphometric information and other attributes.
• The attributes were extracted using traditional object extraction algorithms.
• The method augments traditional image processing.
• Software allowed automatic classification of nuclei of thyroid in microscopic images.

BackgroundThe analyses of several systems for medical-imaging processing typically support the extraction of image attributes, but do not comprise some information that characterizes images. For example, morphometry can be applied to find new information about the visual content of an image. The extension of information may result in knowledge. Subsequently, results of mappings can be applied to recognize exam patterns, thus improving the accuracy of image retrieval and allowing a better interpretation of exam results. Although successfully applied in breast lesion images, the morphometric approach is still poorly explored in thyroid lesions due to the high subjectivity thyroid examinations.ObjectiveThis paper presents a theoretical–practical study, considering Computer Aided Diagnosis (CAD) and Morphometry, to reduce the semantic discontinuity between medical image features and human interpretation of image content.MethodThe proposed method aggregates the content of microscopic images characterized by morphometric information and other image attributes extracted by traditional object extraction algorithms. This method carries out segmentation, feature extraction, image labeling and classification. Morphometric analysis was included as an object extraction method in order to verify the improvement of its accuracy for automatic classification of microscopic images.ResultsTo validate this proposal and verify the utility of morphometric information to characterize thyroid images, a CAD system was created to classify real thyroid image-exams into Papillary Cancer, Goiter and Non-Cancer. Results showed that morphometric information can improve the accuracy and precision of image retrieval and the interpretation of results in computer-aided diagnosis. For example, in the scenario where all the extractors are combined with the morphometric information, the CAD system had its best performance (70% of precision in Papillary cases).ConclusionResults signalized a positive use of morphometric information from images to reduce semantic discontinuity between human interpretation and image characterization.