Inductive Machine Learning and Feature Selection for Knowledge Extraction from Medical Data: Detection of Breast Lesions in MRI

Abstract

This paper presents an approach to the problem of breast cancer diagnosis through the data analysis of magnetic mammography observations (MRi Data), developing corresponding hybrid classification models of patient cases into specific classes (e.g. Benign and Malignant). The aim of this work is the contribution of machine learning to the diagnostic process of breast cancer, offering a supportive intelligent tool that can be used by expert doctors as a medical decision-making aiding tool. Data were collected in collaboration with expert doctors and consist of 77 patient cases. The development of the presented classification models is a combination of inductive decision trees, clustering and feature selection techniques. Specifically, nine (9) different classification models were developed and evaluated by using statistical criteria, medical expert knowledge and where possible, using the Chi-Square statistical test. The performance achieved is considered encouraging for application in real-world practice, while further research is underway for associating MR imaging data with data from invasive examinations (biopsies).

Appendices

Annex 6.1—Abbreviations

Abbreviations	Descriptions
ACR BI-RADS	American College of Radiology Breast Imaging Reporting & Data System
ADC	Apparent Diffusion Coefficient
ANN	Artificial Neural Networks
AUC	Area Under the ROC Curve
B–ANN	Bayesian Artificial Neural Networks
CAD	Computer Aided Diagnostic systems
CHAID	Chi-squared Automatic Interaction Detection
CLi Data	Clinical image Data
CNN	Convolutional Neural Network
DCE–MRI	Dynamic Contrast–Enhanced Magnetic-Resonance Imaging
DMi Data	Digital Mammography image Data
eBP–FNN	error Back Propagation Feed Forward Artificial Neural Networks
FCM	Fuzzy c-Means
FFDM	Full Field Mammograms
FOV	Field of View
HiSS–MRI	High Spatial and Spectral resolution MRI
IRi Data	Infrared Thermography image Data
KNN	K Nearest Neighbor
LDA	Linear Discriminant Analysis
LS–SVM	Least-Squares Support Vector Machine
MARS	Multivariate Adaptive Regression Splines
MCi Data	Microscope image Data
MRi Data	Magnetic–resonance image Data
MSM	Multisurface method
PFK-SVM	Polynomial Kernel Function Support Vector Machine
RF	Random Forest
SA	Sparse Autoencoder
ST	Slice thickness
SVM	Support Vector Machines
TNM	T category describes the primary tumor site, N category describes the regional lymph node involvement and M category describes the presence or otherwise of distant metastatic spread (UNICC- Union for International Cancer Control
TSE	Turbo Spin Echo

Annex 6.2—Variables Frequency Charts (Original Dataset)

Annex 6.3—Variables’ Values Range

Variables’ values range
S. no.	Variable	Values	Group	Frequency (N)	Frequency (%)
1	AGE	[0, 49]	AGE_1	39	50.6
		[50, ∞]	AGE_2	38	50.4
		Total		77	100
2	Morphology (MORPH)	MASS	MORPH_1	48	62.3
		MASS & NON MASS	MORPH_2	2	2.6
		NON MASS	MORPH_3	27	35.1
		Total		77	100
3	Borders (BDS)	IRR	BDS_1	66	85.7
		SPIC
		SMH	BDS_2	11	14.3
		Total		77	100
4	Tumor Size (TUMS)	[0, 1.3]	TUMS_1	28	36.4
		[1.4, 2.5]	TUMS_2	19	24.7
		Over 2.6	TUMS_3	30	39.0
		Total		77	100
5	Peritumoral Edema (PRED)	NO	PRED_1	50	64.9
		YES	PRED_2	27	35.1
		Total		77	100
6	T2 Weighted image (T2-Wi)	NONE
		LOW	T2WI_1	70	90.9
		INTER
		HIGH	T2WI_2	7	9.1
		Total		77	100
7	Curve Morphology (CRM)	TYPE_1	CRM_1	14	18.2
		TYPE_2	CRM_2	25	32.5
		TYPE_3	CRM_3	38	49.4
		Total		77	100
8	Breast Density (BD)	A	BD_1	28	36.4
		B
		B-C	BD _2	49	63.6
		C
		C-D
		Total		77	100
9	Background Parenchymal Enhancement (BPE)	MIN	BPE_1	52	67.5
		MILD
		MOD	BPE_2	25	32.5
		MARK
		Total		77	100
10	Feeding Vessel (FV)	NO	FV_1	46	59.7
		YES	FV_2	31	40.3
		Total		77	100
11	Internal Enhancement (INTEN)	HOMOG	INTEN_1	16	20.8
		INHOMOG	INTEN_2	61	79.2
		HETER
		Total		77	100
12	Diffusion (DF)	N/A	DF_1	4	5.2
		LOW	DF_2	15	19.5
		HIGH	DF_3	58	75.3
		Total		77	100
13	Apparent Diffusion Coefficient (ADC)	N/A	ADC_1	4	5.2
		LOW	ADC_2	48	62.3
		HIGH	ADC_3	25	32.5
		Total		77	100
14	Focality (FC)	U	FC_1	48	62.3
		MC	FC_2	4	5.2
		MF	FC_3	25	32.5
		Total		77	100
15	Benign or Malignant (BOM)	BENIGN	BOM_1	19	24.7
		MALIGNANT	BOM_2	58	75.3
		Total		77	100
16	Benign and Malignant (BAM)	BENIGN	BAM_1	19	24.7
		DCIS	BAM_2	6	7.8
		IDC	BAM_3	36	46.8
		DCIS & IDC	BAM_4	9	11.7
		ILC	BAM_5	6	7.8
		SOLID PAPILLARY	BAM_6	1	1.3
		Total		77	100
17	Malignan (ML)	DCIS	ML_1	6	7.8
		IDC	ML_2	32	41.6
		DCIS & IDC	ML_3	13	16.9
		ILC	ML_4	6	7.8
		SOLID PAPILLARY	ML_5	1	1.3
		Total		58	75.3
		Missing	System	19	24.7
		Total		77	100
18	BIRADS	CATEGORY 3	BIRADS_2	36	46.8
		CATEGORY 4
		CATEGORY 5	BIRADS_3	41	53.2
		CATEGORY 6
		Total		77	100
19	Tumor Grade (TUMG)	GRADE 1	TUMG_1	3	3.9
		GRADE 2	TUMG_2	25	32.5
		GRADE 3	TUMG_3	28	36.4
		Total		56	72.7
		Missing	System	21	27.3
		Total		77	100
20	Estrogen Receptors (ER)	NO	ER_1	14	18.2
		YES	ER_2	43	55.8
		Total		57	74.0
		Missing	System	20	26.0
		Total		77	100
21	Progesterone Receptors) (PR)	NO	PR_1	19	24.7
		YES	PR_2	38	49.4
		Total		57	7 4.0
		Missing	System	20	26.0
		Total		77	100
22	Cerb-B2	NO	CERB2_1	29	37.7
		YES	CERB2_2	28	36.4
		Total		57	74.0
		Missing	System	20	26.0
		Total		77	100
23	Ki-67	[0, 15]	KI67_1	22	28.6
		[16, 25]	KI67_2	8	10.4
		[26, 100]	KI67_3	23	29.9
		Total		53	68.8
		Missing	System	24	31.2
		Total		77	100

Annex 6.4—Classification Tree (Benign or Malignant)