Table 17.

Efficient CNN architectures For medical image understanding

Organ (modality)	Diseases	Tasks	Best architecture
Brain MRI	Alzheimer	Segementation	Lenet-5, GoogleNet
		Detection	Multipath CNN
	Epilepsy	Classification	CNN with seven conv and
	Schizophrenic, bipolar disorder	Segementation	Patch based CNN with 4 conv, 2 maxpool and softmax regression
Breast mammogram	Tumour	Segementation	(Conv + maxpool) $\times$ $2+2$ FC, Softmax
		Detection	Scaled VGGnet or GoogleNet
		Classification	Alexnet
		Localization	Semi Supervised Deep CNN
Heart (CT)	Coronary artery calcium scoring	Segmentation	Two path CNN
Heart (ECG)	ECG	Classification	Two path CNN, one along spatial and the other along temporal and fused together finally
Heart (MRI) left ventricle		Localization	(Conv, maxpool) $\times 2$, FC
Lung (CT)	Cancer	Classification	Ensemble of overfeat across axial, saggital and coronal plane
	Nodule	Detection	Lenet, Overfeat
	COVID-19	Classification	multi scale multi encoder ensemble CNN model
	ILD	Classification	Any CNN architecture like Alexnet
Hep-2 cell		Classification	(Conv, maxpool) $\times 3$, 2 FC, softmax regressor
Eye	Haemorrhage	Detection	(Conv, ReLu, maxpool) $\times 5$, FC, softmax classifier
	Glaucoma	Detection	(Conv, ReLu, maxpool) $\times 2$, FC with adaboost
	Retinopathy	Segmentation	(Conv, ReLu, maxpool) $\times 10$, 3 FC, softmax classifier
Colon	Polyp	Classification	Any simple CNN architecture
	Polyp	Detection	Ensemble of CNN
Skin	Melanoma	Detection	(Conv, ReLu, maxpool) $\times 2$, FC
Liver (CT)		Classification	(Conv, maxpool) $\times 2$, 3 conv, max pool, 3 FC, softmax
Abdomen (US scan)	Fetus	Localization	(Conv, ReLU, maxpool) $\times 2$ conv $\times 3$, maxpool, FC $\times 3$