RAI Labs Duke
Breast CAD | Chest CAD | Lung Dis CT | XCAT-MOBY-ROBY | Breast Models | FE Cardiac Model
Quant. Image | Emerg. Quant. Imaging | Perf. Metrology | Clinical Trials | Emerg. Clinical

Computer-aided diagnosis: Breast imaging

Breast cancer is one of the most devastating and deadly diseases for women today, but early detection leads to significantly better survival rates. However, the diagnostic management of breast cancer is a difficult task. Mammography, the mainstay of breast cancer screening, fails to detect up to 30% of breast lesions and up to 2/3 of those lesions are visible in retrospect. These missed lesions occur despite the fact that physicians err on the side of caution and typically refer all patients with unresolved suspicious findings in their diagnostic mammograms for biopsy. Due to this conservative practice, the majority (60%-80%) of breast biopsies performed in the US turn out to be benign. The problem is further compounded by the large performance gap between specialized breast imagers and general radiologists. Yet, experienced radiologists read less than 10% of mammographic studies.

Computerized decision support systems in mammography provide a fast, reliable, and cost-effective “second opinion” to aid physicians with the diagnostic management of breast cancer. Several computer aided detection (CADe) and computer aided diagnosis (CADx) systems have been developed over the past two decades. Several commercial CADe products are available for mammography and clinically accepted. Our group has been active in both the CADe and CADx research fronts.

Computer-Assisted Detection of Breast Cancer (CADe): The two most commonly detected signs of breast cancer are masses and calcifications, but detecting masses is a far more challenging task for radiologists and computers alike.

We have designed several computer aided detection (CADe) systems to help radiologists with early detection of subtle masses, exploring different strategies ranging from trainable systems to knowledge-based systems; and from feature-based to featureless algorithms. For this research, we rely extensively on DOD’s Digital Database for Screening Mammography (DDSM), as it is the largest publicly available breast cancer database and provides digitized mammograms with proven outcomes. Thus far, our research has yielded the following distinct CADe systems:

1. CADe System based on novel Laguerre-Gauss channelized Hotelling observer (LG-CHO):


This CADe system consists of five steps. First an image processing filter detects suspicious regions that are localized and segmented (Fig. 1: steps A and B). Then the suspicious regions are analyzed by extracting 36 digital image features based on morphology and texture. Extraction of these image features is enhanced by two novel vision-based Hotelling observers that we developed (Fig. 1: step C) called the sub-region Hotelling observer and Laguerre-Gauss channelized Hotelling observer (LG-CHO). Hotelling observer is the optimal linear detector when information about the signal, background, and covariance matrix are known. Feature selection followed by linear discriminant statistical models (Fig. 1: steps D and E) merges the extracted features to predict whether each suspicious region is an actual breast mass or not. Based on up-to-date studies from 1400 DDSM mammograms, the final system delivers up to 90% sensitivity. However, the system’s specificity is lower than that of commercial systems (see example images in Figure 2), which is an area of further research.

 

Figure 1: CADe Flowchart

 

    Figure 2: Example Images    
   
   
77 year old woman with ill-defined, irregular mass (left shows mammogram with radiologist's hand-drawn outline of lesion). CAD detects the mass as well as generating 2 false positives in this image (right).
   
   
   
48 year old woman with ill-defined, irregular mass (left). CAD placed 2 true positive marks in the lesion and generated no false positive marks (right).
   
   
   
38 year old woman with ill-defined, irregular mass (left). CAD placed 2 true positive marks in the lesion and generated no false positive marks (right).
   
     

 

2. CADe System based on Information-Theory (IT-CADe):

For CADe technology to play a significant supporting role in breast cancer screening, it needs to have

(i) adequate stand-alone sensitivity and specificity,
(ii) the ability to capitalize on the knowledge (i.e., new cases) accumulating in digital image libraries,
(iii) synergistic integration with its user to maximize clinical benefit,
(iv) the flexibility to meet the user’s different needs (e.g., on-site evidence-based decision support, telemammography, training and continuing education), and
(v) the ability to adapt to changes in breast imaging technology as it transitions from screen-film mammography (SFM) to full-field digital mammography and also to future techniques such as digital breast tomosynthesis.

Over the past 5 years we have developed a CADe system for mass detection that relies on algorithms better suited to meet the desired conditions listed above. Our system follows a knowledge-based (KB) approach which relates each new case to similar cases stored in its knowledge databank. A diagnosis is assigned to the new case by examining similar knowledge cases. What sets our system apart from other KB systems is that we apply information theory to assess case similarity.

Thus far, our information-theoretic CADe system (IT-CADe) has been evolving as an interactive system designed to offer a "second opinion" for areas noted by a radiologist or marked as suspicious by other CADe schemes. Basically, radiologists review mammograms and select areas that they would like to scrutinize further with IT-CADe. In this interactive capacity, IT-CADe is designed to address mainly the interpretational (not perceptual) error associated with the detection of mammographic masses (Figure 3). Our studies show that IT-CADe could substantially reduce the false positive rate of both existing prescreening systems and expert mammographers while operating at a mass sensitivity rate that is greater than 90% per image.

 

Figure 3: IT-SCAN Prototype

Computer-Aided Diagnosis (CADx) of Breast Cancer:

CADx deals with the application of computer tools to provide recommendations regarding the malignancy status and clinical management of breast lesions detected during mammographic screening. Typically a CADx system is expected to help radiologists determine if an already detected lesion requires biopsy to determine its malignancy status. Reducing the number of unnecessary biopsies is an important clinical task due to the economic cost, physical burden, and emotional stress associated with excessive biopsy of benign lesions. Furthermore, another well-documented problem is the variability among radiologists regarding the recommended clinical management (biopsy vs. short-term follow-up) of suspicious breast lesions. CADx systems aim to improve the sensitivity, specificity, efficiency, and cost-effectiveness of breast cancer screening programs.

 

Fig 4. A Flow Chart showing a typical CADx system.

 

In our laboratory we have developed CADx classifiers that span a wide range of machine learning decision algorithms including artificial neural networks, case-based reasoning, knowledge-based analysis, support vector machines, linear and likelihood ratio-based statistical modeling, and novel decision fusion strategies. Our CADx algorithms are designed to capitalize on various sources of clinical information such as computer-extracted image features, radiologist-interpreted findings, and patient history findings. The features are carefully selected and merged together using a classifier (Fig. 4). These decision models have held up to repeated validation tests using ROC analysis and clinically relevant performance indices such as positive predictive value and partial ROC area even as our databases grew due to newly acquired data from multiple clinical sites,

 

In a recent study on the independent clinical evaluation of our case-based reasoning and likelihood-ratio based CADx systems, both systems were able to correctly identify 98% of malignant masses while sparing almost half of the benign masses from unnecessary biopsy. To enhance these predictive models with interpretability, we are also pursuing content-based image retrieval (CBIR) as the basis of a CADx system. When a CADx system is given a lesion to evaluate, the CBIR algorithm retrieves similar cases from a knowledge databank. The retrieved images are presented to the radiologist in the rank order according to similarity with their corresponding malignancy status. In this manner, radiologists are given visual justification of the computer’s recommendation (Fig. 5).

This work is funded by NIH (PI: Tourassi)

Figure 5: Similarity Matches