Construction of adaptive pulse coupled neural network for abnormality detection in medical images

Figures & data

Figure 1. Representation of biological model.

Figure 2. Mathematical modification of PCNN model.

Figure 3. Architecture of APCNN for lesion detection.

Figure 4. Internal architecture for APCNN.

Figure 5. PCNN harmonics of dermoscopic lesion.

Table 1. Parameters of the proposed PCNN at which various other skin lesions are detected.

Dermoscopic images	FDT	TDT	FDCV	LMCV	TMC	Linking ()
Actin keratosis	0.002	20	0.02	2	10	0.4
Basal cell carcinoma	0.003	10	0.01	2	20	0.3
Dermotofibroma	0.001	10	0.04	1	10	0.3
Haemangioma	0.002	10	0.02	2	10	0.5
Intraepithelial carcinoma	0.002	10	0.02	1	20	0.3
Melacytic nevi	0.002	10	0.03	1	20	0.5
Malignant melanoma	0.002	10	0.02	1	10	0.5
Pyogenic granuloma	0.002	20	0.02	2	10	0.4
Squamous cell carcinoma	0.001	20	0.03	1	20	0.3
Seborrhoeic keratosis	0.004	10	0.03	1	20	0.5

Figure 6. Maximum energy and entropy of pixel at image harmonic.

Table 2. Number of samples for each class of lesion available in dermoscopic standard dataset of dermofit images (Ballerini et al. 2013).

Lesion class	No of samples	Lesion class	Noncancerous/precancerous/cancerous
Actin keratosis	45	Malignant	Precancerous
Basal cell carcinoma	239	Malignant	Cancerous
Dermoto fibroma	65	Benign	Tumor
Haemangioma	97	Benign	Tumor
Intraepithelial carcinoma	78	Malignant	Precancerous
Melacytic nevi	331	Normal	Noncancerous/tumoric
Malignant melanoma	76	Malignant	Cancerous
Pyogenic granuloma	24	Benign	Tumor
Squamous cell carcinoma	88	Malignant	Cancerous
Seborrhoeic keratosis	257	Benign	Tumor

Table 3. Comparison of existing PCNN with APCNN in terms computational speed.

Dermoscopic images size (100 × 100)	PCNN	PCNN	APCNN	APCNN
Actin keratosis	9.559s	96	0.601s	4(11,69)
Basal cell carcinoma	8.037s	97	0.581s	3(42,78)
Dermotofibroma	10.182s	98	1.216s	2(42,78)
Haemangioma	9.734s	96	0.594s	4(9, 83)
Intraepithelial carcinoma	10.117s	98	0.823s	2(16,62)
Melacytic nevi	9.253s	97	2.166s	3(38,99)
Malignant melanoma	15.558s	96	1.540s	4(12,85)
Pyogenic granuloma	7.825s	97	0.835s	3(13,73)
Squamous cell carcinoma	10.55s	97	3.022s	3(10,81)
Seborrhoeic keratosis	15.376s	98	1.836s	2(13,57)

Figure 7. Comparison of Hammond distance.

Figure 8. Comparison of gray level contrast.

Figure 9. Comparison of overall accuracy for lesion detection.

Figure 10. Entropic threshold obtained in APCNN and PCNN.

Table 4. Accuracy comparison between APCNN and existing PCNN (Johnson and Padgett 1999; Wise 2007) model for binary classification of each lesion class.

Dermoscopic images	APCNN			PCNN
Lesion class	Precision	Recall	Accuracy	Precision	Recall	Accuracy
Actin keratosis	1.0000	0.9738	0.9735	1.0000	0.9728	0.9632
Basal cell carcinoma	1.0000	0.9675	0.9662	1.0000	0.9652	0.9657
Dermotofibroma	0.9943	1.0000	0.9964	0.9933	0.9999	0.9934
Haemangioma	0.9997	0.9995	0.9983	0.9987	0.9992	0.9981
Intraepithelial carcinoma	1.0000	0.9997	0.9556	1.0000	0.9987	0.9536
Melacytic nevi	0.9678	1.0000	0.9794	0.9676	0.9999	0.9784
Malignant melanoma	0.9556	1.0000	0.9844	0.9546	0.9999	0.9834
Pyogenic granuloma	0.9802	0.9999	0.9739	0.9800	0.9996	0.9738
Squamous cell carcinoma	0.9713	0.8045	0.8124	0.9703	0.8025	0.8014
Seborrhoeic keratosis	1.0000	0.8766	0.8768	1.0000	0.8756	0.8766

Ballerini, L., et al. 2013. A color and texture based hierarchical K-NN approach to the classification of non-melanoma skin lesions. In: Celebi M., Schaefer G. (eds.), Color Medical Image Analysis. Lecture Notes in Computational Vision and Biomechanics, vol 6. Springer.

 Google Scholar

Johnson, J. L., and M. L. Padgett. 1999. PCNN models and applications. IEEE Transactions Neural Networks 10 (3):480–498. doi:10.1109/72.761706.

 PubMedGoogle Scholar

Wise, R. 2007. Optoelectronic implementations of pulse coupled neural network: Challenges and limitations. Cambridge, MA: Massachusetts Institute of Technology (MIT).

 Google Scholar