Automated Cancer Detection and Tissue Classification

 

This portfolio focuses on the automated detection of two prevalent types of cancer: breast and prostate. Both types of cancer are very prevalent, impacting many individuals per year. Manually-conducted histological assessment of potentially cancerous tissue by a highly trained pathologist forms the current gold standard of diagnosis. Several samples are usually taken and several sections from each section are examined: roughly 20 to 120 sections are examined per patient frequently by more than one pathologist. Such manual assessment suffers from several limitations, the most significant being inter-pathologist variation. Exact intra-pathologist agreement has been shown to vary widely (i.e. 43-78%) thereby leading routinely to variability in disease grading and an intense time requirement to effectively examine each and every section.

Automated pathological assessment of tissue sections provides an attractive solution to the problems of significant assessment time and inter-pathologist variability in diagnosis. This portfolios automated technologies focus on the tissue vasculature, lumens, and the epithelial linings as this is where cancer presents most identifiably.

Single Modality Automated Analysis

This technology exploits the properties of Fourier Transform Infrared Spectroscopy (FT-IR) to create direct visualizations of both tissue structure and chemistry thereby allowing a fine automation process based on easily quantifiable structures and trends.

Here, many tissue samples are analyzed using tissue microarrays and FT-IR. The resulting spectra are analyzed and both the spectral and spatial data are used to first segment the images into epithelial and stromal content. Following epithelial segmentation, several spectral metrics are quantified in concert with the spatial data for each tissue sample and the epithelial segmented images are further segmented into cancerous and normal epithelial content: epithelial pixels are separated into cancerous and normal classes based on epithelium content and organization. All segmentation is done using proprietary software designed and written by UIUC researchers specifically for this application. Our technique of using FT-IR for epithelial/stromal and subsequent cancerous/normal segmentation via chemical fingerprint and properties shows excellent promise: ROC analysis shows a nearly perfect discrimination between normal and cancerous tissue samples.

Multimodal Automated Analysis

Here, multimodal images (optical and spectroscopic) are created and analyzed for epithelial and lumenal irregularities. Normal tissue generally has smooth lumens with thin epithelial layers whereas cancerous tissue generally has smaller elliptical or circular lumens with larger or thicker epithelial layers. As malignancy advances, the average size of individual epithelial cells and their nuclei tends to increase. The quantification algorithm created and used by UIUC researchers automatically detects tissue abnormalities traditionally indicative of cancerous tissue. Additionally, the software is capable of quantifying the malignancy, or stage, of a cancer based on its physical features.

Tissue samples are initially stained and imaged using optical microscopy then classified using FT-IR. The optical and spectroscopic images are scaled properly, registered and overlaid to create a multi-modal image containing a wealth of information. Cancerous tissue has many identifying factors; however, epithelial morphology is one of the most identifiable and is the clinical gold standard. Therefore, this automated technique analyzes this factor (epithelial morphologycellular and nucleic, as well as lumen geometry) as its primary diagnostic source. The image is then thoroughly segmented through several techniques including, representatively: lumen detection, nucleus detection, tissue segmentation (between stromal and epithelial), etc. After segmentation and feature detection, epithelial features are analyzed globally and locally to create a diagnosis recommendation. Combining FT-IR and optical techniques was found to synergistically produce excellent accuracy allowing robust, reproducible cancer cell detection, and tangentially, detailed automated general tissue segmentation.

Automated Tissue Referencing

This addition to automated cancer detection and classification does not independently arrive at a diagnosis but rather reduces the time for a pathologist to arrive at his own conclusion. Additionally, it can also be used to improve the accuracy of other automated systems. This computer information, management, and decision-making system relies on one or more characteristic features of an input tissue image to provide reference images from a preexisting database that are similar to the sample under consideration. Ultimately, this software provides the closest matching cases to an input tissue, as well as their diagnosis, thereby allowing a physician to compare samples quickly and efficiently.

Tissue sample images are loaded into proprietary software and initially morphological features are extracted from the unknown sample. Next, similarities between the uploaded tissue sample(s) and the reference samples are calculated based on commonality of morphological features. Lastly, the most similar reference tissue samples are retrieved from the database for pathologist review (or automated comparison in a fully automated method). Tissue samples with like characteristics and patterns as the sample of interest will afford significant information to pathologists during diagnosis thereby speeding diagnosis time and decreasing pathologist variability.

Applications

This portfolio is targeted at providing a sensitive, accurate means of automated detection and classification of cancerous vs. healthy cells in tissue samples. This will allow significant, direct benefits to pathologists and patients, including decreased workload and improved diagnosis variability respectively, and indirect, runoff benefits such as lowered costs.

Beyond cancer detection, however, this portfolio can be easily extended to detect other types of cells and tissues.

Benefits

Single Modality Automated Analysis

  • Excellent Immunity to Sample Variation: FT-IR provides a very robust solution to automated cancer detection because it accurately and sensitively evaluates chemical content; hence, this technique is unaffected by significant impediments to current pathological assessment, such as application of contrast agents, or the use of chemical stains.
  • Predominant Automation: This technique is largely automated, only requiring proper sample preparation and loading. Removing the downstream human requirements negates a considerable source of error: a team of pathologists analyzing thousands of samples yearly has ample opportunities for error. Our automated system removes the possibility of error derived from such common sources as sample misplacement between pathologist review as well as the more difficult problem of intra-pathologist variability.

Multimodal Automated Analysis

This technology includes FT-IR as well and includes the benefits from the above technology as well.

  • FT-IR and Optical Combination: Our unique, synergistic combination of tissue spectroscopic and common stained optical images allows very stable, robust, and accurately reproducible results for automated cancer detection.
  • Algorithm Accuracy: The proprietary software written by UIUC researchers for this application has been shown to have excellent accuracy and is currently ready for translation to clinical settings.

Automated Tissue Referencing

  • Increased Accuracy: This technology provides a significant increase in accuracy when combined with any of the other techniques in this portfolio or when used by a human pathologist operator. It effectively pulls similar, already-diagnosed, ground-truth images from a database allowing quicker and more accurate diagnosis.