computational intelligence for the detection and classification of malignant lesions in screening...
TRANSCRIPT
COMPUTATIONAL INTELLIGENCE FOR THE
DETECTION AND CLASSIFICATION OF
MALIGNANT LESIONS IN SCREENING MAMMOGRAPHY
DATA
E. Panourgias, A. Tsakonas, G Dounias, G. Panagi (Athens, Thessaloniki, Chios, Greece)
E-mail: [email protected]
Lancet 2003 Vol 361,Apr 26
• Early detection and diagnosis of breast cancer represents a very important factor in its treatment and consequently the survival rate
• Screening mammography is considered the most reliable method of early detection, accounting for a decrease in mortality of up to 18-23%
Long term effect of screening mammmography on breast cancer death in 2 Swedish counties
20
40
60
80
100
120
140
160
1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004
45-49
50-59
60-69
70-74
Age group
45-49
50-59
60-69
70-74
Age group Source: LETB
Mo
rtal
ity
rate
per
100
,00
0
Year
23%
29%
18%
LANCET 2003 VOL 361, APR 26
Mammographic Appearance of Breast Cancer
• Spiculated masses
• Pleiomorphic,Heterogeneous Microcalcifications
• Focal asymmetric densities with ill-defined margins or microlobulations
• Architectural distortion
Mammographic Appearance of Breast Cancer
• Spiculated masses
• Pleiomorphic,Heterogeneous Microcalcifications
• Focal asymmetric densities with ill-defined margins or microlobulations
• Architectural distortion
AIM OF STUDY
• We used data of 200 histologically proven malignant lesions discovered during screening to develop computer algorithms that may point in the direction of a specific histologic diagnosis.
• Machine learning and Genetic Programming were applied.
INDUCTIVE MACHINE LEARNING
• Method of computational intelligence based analysis
• Has the ability to process large and complex databases
• Constructs decision trees by intelligently reducing either,
• Complexity of the search space or the size of the tree.
GENETIC PROGRAMMING
• Operates by mimicking a living population
• Survival of the fittest (fitness is how successful a member is in completing its assigned task- the least fit members are eliminated )
• New members added (mutation, breeding, random generation)
- a population of random programs is generated
MATERIALS AND METHODS • For each case, all 4 standard views
were used, as well as clinical and pathology data
• All cases were rated according to the level of concern by using standard Breast Imaging Reporting and Data System, or BIRADS, recommendations
BIRADS LEXICON
CATEGORY 1 NEGATIVE
CATEGORY 2 BENIGN FINDING
CATEGORY 3 PROBABLY BENIGN-MALIGNANCY CANNOT BE EXCLUDED
CATEGORY 4 SUSPICIOUS ABNORMALITY-BIOPSY RECOMMENDED
CATEGORY 5 HIGHLY SUGGESTIVE OF MALIGNANCY
ATTRIBUTES • Age
• Mammographic parenchymal pattern (Pattern 1-5)
• Rt-Lt breast, Position-quadrant: Upper outer, upper inner, lower outer, lower inner, retroareolar
• Mass-shape, margins
• Microcalcifications
• Architectural distortion
ATTRIBUTES
• Associated findings (nipple retraction, skin thickening)
• BIRADS score
• Histologic diagnosis
• Histologic size
• Lymph node status
• Estrogen Receptor status
• Progesterone Receptor status
MAIN HISTOLOGIC TYPES OF BREAST CANCER
• Ductal cancer
DCIS (ductal carcinoma in situ)
Invasive ductal carcinoma
• Lobular carcinoma
LCIS (lobular carcinoma in situ)
Invasive lobular carcinoma
Ductal Carcinoma
• Over 80% are variants of ductal carcinoma
• Two types:• Noninvasive (ductal carcinoma in situ-
DCIS): tumor cells are confined to the duct epithelium and do not penetrate the basement membrane
• Invasive (IDC) tumor cells penetrate the basement epithelium and invade the surrounding tissues
Lobular Carcinoma• Noninvasive type or lobular carcinoma in situ
(LCIS)Does not form a palpable mass or visible lesion by mammographyCurrently classified as a PREMALIGNANT lesion rather than a true cancer
• Invasive Lobular Carcinoma (ILC)Tends to be bilateral more often than ductal carcinoma (20% of cases are bilateral)Tend to be multicentric within the same breast
Extracted Rule 1
If a mass with ill-defined margins, is
observed in the RT breast in the
UOQ, it is most likely IDC
Statistical prediction (0.875)
The presence of an ill-defined or
spiculated mass on a mammogram
is almost pathognomonic of an
Invasive Ductal Carcinoma
D. Kopans, Breast Imaging, 2nd ed., 1998
Extracted Rule 2
If patient presents with a Focal Asymmetric Density and a BIRADS score 3 in the RT breast, lesion
is suggestive of invasive ductal carcinoma (IDC) if size is <14mm
and invasive lobular carcinoma (ILC) if the lesion is >14mm or in BOTH breasts
(0.867)
ILC cells have decreased E-cadherin
expression which is a glue-like
substance that provides cell-to-cell
adhesion, a feature prominent in
IDC that causes cells to stick
together and produce a
mammographically visible mass
Neal Goldstein Am J Clin Pathology118(3):425-434,2002
• This is why ILC is frequently less apparent on mammograms and therefore, generally larger at diagnosis
• Silverstein et al found that the average size at diagnosis for IDC’s was 23mm and for ILC’s 30mm.
Cancer 1994;73:1673-1677
• ILC tends to be bilateral more often than ductal carcinoma (20% of cases are bilateral)
Decision Tree Results
• Woman presenting with a suspicious lesion and BIRADS score 5
in the UIQ of the RT breast and
size of lesion is <21mm, then it is IDC, >21mm it is ILC
• Woman presenting with a FAD with BIRADS score 5
lesion size of <42mm then it is IDC,
>42mm, it is ILC
In a study that included 50 000 IDC’s and ILC’s, Arpino et al found that
• 54% of ILC’s are larger than 2cm, compared to 48% of IDC’s
• 14% of ILC’s presented as a large tumor exceeding 5cm, as compared with 9% of IDC’s
Grazia Arpino et al. Breast cancer Res 2004;6(3)R149-156.
Decision Tree Results
If patient with a BIRADS score 4
presents with suspicious microcalcifications
(MC) on a mammogram in the UOQ and
an associated Architectural Distortion (AD),
then she is more likely to have IDC,
whereas if the MC are not accompanied by
AD, then the diagnosis of DCIS is more
probable
• DCIS is a form of malignant transformation of the epithelial cells lining the mammary ducts and lobules
• The proliferating cells are confined by an intact basement membrane
• Necrotic debris in the lumen of the duct produces microcalcifications which are visible on a mammogram
Extracted Rules of GP
• If the mass margin is equal to or greater than 3, then the histology diagnosis is IDC
• If the mass margin is < 3 and the size <1cm, then the lesion is IDC, if it is > 1cm it is ILC.
Values of variables- Mass margin: 0=circumscribed, 1=ill-defined, 2= lobulated, 3=obscured, 4=spiculated
ConclusionsDespite the limited information (no prior studies, no normal cases, many more cases of IDC than other types of cancer) and the fact that different types of abnormalities (MC, masses, AD) were included , the classification performances of determining that an identified lesion was a specific histological subtype was reasonable and consistent
Conclusions
• The extracted rules often included the RT breast as a determining factor- needs further evaluation as this has not been proven in the literature
• the computerized classification methods often used histology findings such as size to categorize the mammographic lesions
Conclusions
These issues have to be further investigated with larger datasets that include a greater number of attributes, a substantial amount of normal patients and more cases of cancers other than IDC’s, that composed that majority of our present dataset