06-03-2013, 04:24 PM
Automated Method for Improving System Performance of Computer-Aided Diagnosis in Breast Ultrasound
Automated Method for Improving.pdf (Size: 862.84 KB / Downloads: 36)
Abstract
The purpose of this research was to demonstrate the
feasibility of a computerized auto-assessment method in which a
computer-aided diagnosis (CADx) system itself provides a level
of confidence for its estimate for the probability of malignancy
for each radiologist-identified lesion. The computer performance
was assessed within a leave-one-case-out protocol using a database
of sonographic images from 542 patients (19% cancer
prevalence). We investigated the potential of computer-derived
confidence levels both as 1) an output aid to radiologists and 2)
as an automated method to improve the computer classification
performance—in the task of differentiating between cancerous
and benign lesions for the entire database. For the former, the
CADx classification performance was assessed within ranges of
confidence levels. For the latter, the computer-derived confidence
levels were used in the determination of the computer-estimated
probability of malignancy for each actual lesion based
on probabilities obtained from different views.
INTRODUCTION
COMPUTER-AIDED diagnosis (CADx) has the potential
to be a valuable decision making tool for radiologists.
Generally, CADx involves the computerized estimation of the
probability of malignancy of an imaged lesion and could help a
radiologist in recommending patient work-up strategy (lesion
biopsy versus imaging follow-up). However, CADx output is
only helpful if radiologists understand the capabilities and limitations
of the computer analysis. The many automated steps involved
in a CADx scheme—lesion segmentation, feature extraction,
and classification for malignancy—are inherently complicated
and hence each prone to a certain degree of error. Therefore,
“uneducated” use of CADx in a diagnostic setting could
negatively influence diagnosis, e.g., lead to more unnecessary
biopsies or to more misclassified cancers. In addition, radiologists
could lose confidence inCADx when presented withCADx
results that are obviously incorrect. This, in turn, could induce
radiologists to ignore CADx output in most instances—if not to
completely abandon CADx—when in fact CADx output might
be helpful.
MATERIALS AND METHODS
Patients and Lesions
The database consisted of consecutive diagnostic breast ultrasound
examinations collected under protocols approved by the
Institutional Review Board and HIPAA constraints. There was
no other case selection criterion for inclusion in this study. Informed
consent was obtained from 542 patients and their sonographic
images were used in this study. These patients presented
with a total of 1133 distinct abnormalities that were each imaged
in one or more views (Table I). It is our standard clinical
practice to save two orthogonal views of each lesion and obtain
additional views at the discretion of the attending radiologist.
All abnormalities noted in the radiology reports were included
in our analysis whether or not the pathology was proven
by biopsy. Pathology was proven by biopsy in 32% of the cases
(363/1133). The clinical positive predictive value for biopsywas
44% (158/363). There were 105 breast cancer patients with a
total of 158 cancerous lesions bringing the cancer prevalence in
this study population to 19% by patient (105/542) and to 14%
by lesion (158/1133). The most prevalent lesion type was cystic
with the majority being small subcentimeter cysts.
CADx With Confidence Levels
The metric used to determine the confidence level of a
given computer-extracted lesion contour was the overlap of
the computer-segmented lesion area with the area delineated
by the radiologist. The overlap of two regions was defined as
the area of their intersection divided by the area of their union.
Hence, it was a normalized entity and will be referred to as
“actual overlap” in this paper. It is important to note that the
“actual overlap” did not depend strongly on fine details of
the manually-drawn contours. The first step in the calculation
of auto-assessment confidence level (AACL) values corresponding
to the computer-extracted lesion contours was the
determination of lesion descriptors useful for the task of distinguishing
between poorly-computer-segmented lesions and
those segmented successfully by the computer.
Performance Analysis
The training and testing of the “classification BNN” was
done within a leave-one-case-out protocol and classification
performance was assessed using receiver operator characteristic
(ROC) analysis [11]–[13]. A single leave-one-case-out
analysis was performed in which the four previously described
lesion features for each computer-derived lesion outline formed
the input to the “classification BNN” and the output was
the computer-estimated probability of malignancy for each
lesion outline. In order to obtain an estimated probability of
malignancy for each lesion.