DSNN: A DenseNet-Based SNN for Explainable Brain Disease Classification

ORIGINAL RESEARCH published: 26 May 2022 doi: 10.3389/ nsys.2022.838822 Edi ed by: Robe as Damase icius, Silesian Uni e si y o Technology, Poland Re iewed by: A shin Shoeibi, K.N.Toosi Uni e si y o Technology, I an Dela am Sadeghi, Islamic Azad Uni e si y o Mashhad, I an Ri ys Moskoliunas, Vy au as Magnus Uni e si y, Li huania *Co espondence: Shui-Hua Wang [email p o ec ed] Juan Manuel Go iz [email p o ec ed] Yu-Dong Zhang [email p o ec ed] †These au ho s ha e con ibu ed equally o his wo k Recei ed: 18 Decembe 2021 Accep ed: 25 Ap il 2022 Published: 26 May 2022 Ci a ion: Zhu Z, Lu S, Wang S-H, Go iz JM and Zhang Y-D (2022) DSNN: A DenseNe -Based SNN o Explainable B ain Disease Classi ica ion. F on . Sys . Neu osci. 16:838822. doi: 10.3389/ nsys.2022.838822 DSNN: A DenseNe -Based SNN o Explainable B ain Disease Classi ica ion Ziquan Zhu1†,Siyuan Lu1†,Shui-Hua Wang1,2*, Juan Manuel Go iz3* and Yu-Dong Zhang1,2,4* 1School o Compu ing and Ma hema ical Sciences, Uni e si y o Leices e , Eas Midlands, Uni ed Kingdom, 2School o Compu e Science and Technology, Henan Poly echnic Uni e si y, Jiaozuo, China, 3Depa men o Signal Theo y, Ne wo king and Communica ions, Uni e si y o G anada, G anada, Spain, 4Guangxi Key Labo a o y o T us ed So wa e, Guilin Uni e si y o Elec onic Technology, Guilin, China Aims: B ain diseases e e o in ac anial issue and o gan in lamma ion, ascula diseases, umo s, degene a ion, mal o ma ions, gene ic diseases, immune diseases, nu i ional and me abolic diseases, poisoning, auma, pa asi ic diseases, e c. Taking Alzheime ’s disease (AD) as an example, he numbe o pa ien s d ama ically inc eases in de eloped coun ies. By 2025, he numbe o elde ly pa ien s wi h AD aged 65 and o e will each 7.1 million, an inc ease o nea ly 29% o e he 5.5 million pa ien s o he same age in 2018. Unless medical b eak h oughs a e made, AD pa ien s may inc ease om 5.5 million o 13.8 million by 2050, almos h ee imes he o iginal. Resea che s ha e ocused on de eloping complex machine lea ning (ML) algo i hms, i.e., con olu ional neu al ne wo ks (CNNs), con aining millions o pa ame e s. Howe e , CNN models need many aining samples. A small numbe o aining samples in CNN models may lead o o e i ing p oblems. Wi h he con inuous esea ch o CNN, o he ne wo ks ha e been p oposed, such as andomized neu al ne wo ks (RNNs). Schmid neu al ne wo k (SNN), andom ec o unc ional link (RVFL), and ex eme lea ning machine (ELM) a e h ee ypes o RNNs. Me hods: We p opose h ee no el models o classi y b ain diseases o cope wi h hese p oblems. The p oposed models a e DenseNe -based SNN (DSNN), DenseNe -based RVFL (DRVFL), and DenseNe -based ELM (DELM). The backbone o he h ee p oposed models is he p e- ained “cus omize” DenseNe . The modi ied DenseNe is ine- uned on he empi ical da ase . Finally, he las i e laye s o he ine- uned DenseNe a e subs i u ed by SNN, ELM, and RVFL, espec i ely. Resul s: O e all, he DSNN ge s he bes pe o mance among he h ee p oposed models in classi ica ion pe o mance. We e alua e he p oposed DSNN by i e- old c oss- alida ion. The accu acy, sensi i i y, speci ici y, p ecision, and F1-sco e o he p oposed DSNN on he es se a e 98.46% ±2.05%, 100.00% ±0.00%, 85.00% ±20.00%, 98.36% ±2.17%, and 99.16% ±1.11%, espec i ely. The p oposed DSNN is compa ed wi h es ic ed DenseNe , spiking neu al ne wo k, and o he s a e-o - he-a me hods. Finally, ou model ob ains he bes esul s among all models. Conclusions: DSNN is an e ec i e model o classi ying b ain diseases. Keywo ds: b ain diseases, con olu ional neu al ne wo k, andomized neu al ne wo k, DenseNe , MRI F on ie s in Sys ems Neu oscience | www. on ie sin.o g 1May 2022 | Volume 16 | A icle 838822 Zhu e al. DSNN o B ain Disease Classi ica ion INTRODUCTION B ain diseases e e o in ac anial issue and o gan in lamma ion, ascula diseases, umo s, degene a ion, mal o ma ions, gene ic diseases, immune diseases, nu i ional and me abolic diseases, poisoning, auma, pa asi ic diseases, e c. B ain diseases o en show diso de s o consciousness, sensa ion, mo emen , o au onomic ne e dys unc ion. The e may also be e e , headache, omi ing, and o he men al symp oms. Taking Alzheime ’s disease (AD) as an example, he numbe o pa ien s d ama ically inc eases in de eloped coun ies. By 2025, he numbe o elde ly pa ien s wi h AD aged 65 and o e will each 7.1 million, inc easing nea ly 29% o e he 5.5 million pa ien s o he same age in 2018 (Lynch, 2018). Unless medical b eak h oughs a e made, he numbe o Alzheime ’s pa ien s aged 65 and o e may inc ease om 5.5 million o 13.8 million by 2050, almos h ee imes he o iginal. Now, b ain diseases a e mainly diagnosed by doc o s. Howe e , he manual diagnosis equi es much ime. A he same ime, di e en doc o s may ha e di e en iews on he same examina ion esul s, which has b ough a lo o ouble o pa ien s. Mo e and mo e esea che s use compu a ional me hods (Wang e al., 2021) o classi y b ain diseases. No een e al. (2020) in oduced a mul i-le el me hod using wo DensNe 201 and Incep ion- 3 o diagnose ea ly b ain umo s. Finally, he accu acy o Incep ion- 3 and DensNe 201 we e 99.34% and 99.51%, espec i ely. Amin e al. (2019a) p esen ed a model using magne ic esonance images o au oma ically classi y b ain umo s acco ding o he LSTM model me hod. Wha ’s mo e, his me hod ob ained 0.97 DSC in p ac ical applica ion. Amin e al. (2019b) used a deep lea ning model o p edic heal hy and unheal hy b ain umo slices. A unkuma e al. (2020) in oduced a new model o iden i y ROI loca ion based on b ain umo MRI. The me hod inally go 89% sensi i i y, 92.14% accu acy, and 94% speci ici y. Pu usho am Gumas e and Bai agi (2020) p oposed an algo i hm o ex ac le and igh b ain ea u es. This a icle also in oduced di e en s a is ical ea u e ex ac ion me hods and used a suppo ec o machine o ex ac umo egions om s a is ical ea u es. Cha e jee and Das (2019) p oposed a no el me hod o he segmen a ion o b ain images, which we e di ided in o wo ca ego ies: benign (low le el) and e il (high le el). Bhano hu e al. (2020) p esen ed a new me hod acco ding o R-CNN o de ec umo s and ma k hei loca ion. Finally, he de ec ion and classi ica ion accu acy o he h ee ypes o b ain umo s we e 89.45%, 68.18%, and 75.18%. Na eka e al. (2020) compa ed a ious echnologies o b ain umo segmen a ion models and isualized he in e nal concep s o ha e a deepe unde s anding o how hese echnologies segmen ed wi h high accu acy. Aboelenein e al. (2020) in oduced a no el ne wo k (HTTU-Ne ) o b ain umo cu ing. Huang e al. (2020) p esen ed he di e en ial ea u e neu al ne wo k (DFNN) me hod. The me hod in oduced DFM blocks and combined SE blocks. When he DFM block was in oduced, he accu acy o he wo da abases was imp o ed by 1.8% and 1.3%, espec i ely. Hu and Razmjooy (2020) p oposed a me a heu is ic- based sys em o de ec umo s. Sadad e al. (2021) in oduced a no el model acco ding o UNET a chi ec u e and ResNe 50 as he backbone o he de ec ion o b ain umo s. Kalaisel i e al. (2020) p oposed a pa ch-based-upda ed un-leng h egion g ow h (PR2G) me hod o de ec and segmen umo s. The accu acy o his me hod was 97%. Kaplan e al. (2020) used wo me hods o classi y he h ee di e en ypes o b ain umo s. The wo me hods we e nLBP and αLBP. The highes classi ica ion accu acy o b ain umo s was 95.56%. Khalil e al. (2020) p oposed a new me hod (DA clus e ing) o imp o e he accu acy o ex ac ing ini ial con ou poin s o de ec h ee-dimensional magne ic esonance b ain umo s be e . Khan e al. (2020) p oposed a new me hod, pa ial ee (PART), o de ec b ain umo s o g ade I o g ade IV b ain umo s. This me hod used he ule lea ne o an ad anced ea u e se . Ma and Zhang (2021) p oposed a me hod o in elligen ly de ec b ain umo s based on a ligh weigh neu al ne wo k. Hollon e al. (2020) in oduced a new me hod o he au oma ic de ec ion o b ain umo s by combining SRH 5–7, CNN, and he label- ee op ical imaging me hod. Saba e al. (2020) used a new me hod o de ec b ain umo s. The G asp cu me hod was used o segmen b ain umo symp oms, and VGG-19 was used o ob ain ea u es. Sha i e al. (2020) p oposed an unsupe ised uzzy se me hod o b ain umo segmen a ion. The iangula uzzy median il e enhanced he image o be e de ec b ain umo s. Xu e al. (2020) p esen ed a new s uc u e o he ea ly de ec ion o b ain umo s. The new s uc u e was mainly composed o i e pa s: umo segmen a ion, mo phology, denoising, ea u e ex ac ion, and classi ica ion. Heman h e al. (2011) in oduced a no el me hod (HSBPN) o segmen MR b ain umo images. Naye e al. (2013) in oduced a no el s uc u e o he classi ica ion o he MRI da ase . Chen e al. (2017) p esen ed an imp o ed me hod o de ec ing pa hological b ains. A new classi ie was used in he imp o ed me hod. Shoeibi e al. (2020) inished a e iew on he segmen a ion o he Co id-19 by DL. Shoeibi e al. (2021a) pe o med a comp ehensi e su ey abou he applica ion o DL in he de ec ion o mul iple scle osis. Sadeghi e al. (2021) showed a su ey on he au oma ic diagnosis o he SZ by AI. Shoeibi e al. (2021b) comple ed a comp ehensi e e iew on he applica ion o he a ious AI echniques in he diagnosis o epilep ic seizu es. Shoeibi e al. (2021c) comple ed a e iew o a ious me hods based on DL o au oma ic diagnosis o SZ by elec oencephalog am (EEG) signals. Shoeibi e al. (2022) p oposed a new model o au oma ically de ec Epilep ic seizu es. The p oposed model was based on he DL and he uzzy heo y. Odusami e al. (2022) p oposed a me hod o he ecogni ion o AD. They es ed wo CNN models (DenseNe 201 and ResNe 18) o pe o m his ask. This me hod ob ained 98.86% accu acy, 98.94% p ecision, and 98.89% ecall. Razzak e al. (2022) in oduced a new ne wo k (Pa ialNe ) o de ec AD based on MRIs. This ne wo k achie ed imp o emen s on he AD de ec ion. Ash a e al. (2021) expe imen ed wi h di e en CNN models o de ec AD based on ans e lea ning. Finally, he ine- uned DenseNe go he highes accu acy (99.05%). I b ain diseases a e diagnosed manually, doc o s need o spend a lo o ime on examina ion. Some imes we may encoun e he p oblem ha di e en doc o s ha e di e en iews on he examina ion esul s o he same pa ien . As shown in F on ie s in Sys ems Neu oscience | www. on ie sin.o g 2May 2022 | Volume 16 | A icle 838822 Zhu e al. DSNN o B ain Disease Classi ica ion TABLE 1 | Con ibu ions o s a e-o - he-a me hods. Me hod Con ibu ion No een e al. (2020) A mul i-le el me hod using wo DensNe 201 and Incep ion- 3 was p oposed o diagnose ea ly b ain umo s. Amin e al. (2019a) A model acco ding o he LSTM model me hod using magne ic esonance images was in oduced o classi y b ain umo s au oma ically. Amin e al. (2019b) A deep lea ning model was used o p edic heal hy and unheal hy b ain umo slices. A unkuma e al. (2020) A new model was in oduced o ain MRI b ain umo s o iden i y ROI loca ion. Pu usho am Gumas e and Bai agi (2020) An algo i hm was p oposed o ex ac le and igh b ain ea u es. This a icle also in oduced di e en s a is ical ea u e ex ac ion me hods and used a Suppo Vec o Machine o ex ac umo egions om s a is ical ea u es. Cha e jee and Das (2019) A no el me hod was p oposed o he segmen a ion o b ain images. Bhano hu e al. (2020) A new me hod based on R-CNN was p esen ed o de ec umo s and ma k hei loca ion. Na eka e al. (2020) Va ious echnologies we e compa ed o b ain umo segmen a ion models. Aboelenein e al. (2020) The HTTU-Ne was p oposed o b ain umo cu ing. Huang e al. (2020) The DFNN was p oposed. The me hod in oduced DFM blocks and combined SE blocks. Hu and Razmjooy (2020) A me a heu is ic-based sys em was p esen ed o de ec umo s. Sadad e al. (2021) A no el model acco ding o UNET a chi ec u e and ResNe 50 as he backbone was p oposed o he de ec ion o b ain umo s. Kalaisel i e al. (2020) The PR2G was p oposed o de ec and segmen umo s. Kaplan e al. (2020) Then LBP and αLBP we e used o classi y he h ee di e en ypes o b ain umo s. Khalil e al. (2020) The DA clus e ing was p oposed o imp o e he accu acy o ex ac ing ini ial con ou poin s o de ec h ee-dimensional magne ic esonance b ain umo s be e . Khan e al. (2020) The PART was in oduced o de ec b ain umo s o g ade I o g ade IV b ain umo s. Ma and Zhang (2021) A me hod was p oposed o in elligen ly de ec b ain umo s based on a ligh weigh neu al ne wo k. Hollon e al. (2020) A new me hod was p oposed o he au oma ic de ec ion o b ain umo s by combining SRH 5–7, CNN, and he label- ee op ical imaging me hod. Saba e al. (2020) A new me hod was p oposed o de ec b ain umo s. The G asp cu me hod was used o segmen b ain umo symp oms, and VGG-19 was used o ob ain ea u es. Sha i e al. (2020) An unsupe ised uzzy se me hod was in oduced o b ain umo segmen a ion. Xu e al. (2020) A new s uc u e was p oposed o he ea ly de ec ion o b ain umo s. The new s uc u e was mainly composed o i e pa s: umo segmen a ion, mo phology, denoising, ea u e ex ac ion, and classi ica ion. Heman h e al. (2011) The HSBPN was p oposed o segmen MR b ain umo images. Naye e al. (2013) A no el s uc u e was p esen ed o he classi ica ion o he MRI da ase . Chen e al. (2017) An imp o ed me hod was in oduced o de ec ing pa hological b ains. Shoeibi e al. (2020) A e iew was p esen ed on he segmen a ion o he Co id-19 by DL. Shoeibi e al. (2021a) A comp ehensi e su ey abou he applica ion o DL in he de ec ion o Mul iple Scle osis Sadeghi e al. (2021) A su ey was p esen ed on he au oma ic diagnosis o he SZ by AI. Shoeibi e al. (2021b) A comp ehensi e e iew was p esen ed on applying he a ious AI echniques in he diagnosis o Epilep ic seizu es. Shoeibi e al. (2021c) A e iew o a ious me hods based on DL o au oma ic diagnosis o SZ by elec oencephalog am (EEG) signals was comple ed. Shoeibi e al. (2022) A new model was p oposed o de ec Epilep ic seizu es au oma ically. The p oposed model was based on he DL and he uzzy heo y. Odusami e al. (2022) A me hod was p oposed o he ecogni ion o AD. They es ed wo CNN models (DenseNe 201 and ResNe 18) o pe o m his ask. Razzak e al. (2022) A new ne wo k (Pa ialNe ) was in oduced o de ec AD based on MRIs. This ne wo k achie ed imp o emen s in AD de ec ion. Ash a e al. (2021) Di e en CNN models we e expe imen ed wi h o de ec AD based on ans e lea ning. Finally, he ine- uned DenseNe go he highes accu acy (99.05%). Table 1, mos esea che s use deep con olu ion neu al ne wo ks (DCNNs) o classi y and iden i y b ain diseases. Howe e , he e will be many pa ame e s and calcula ions in he aining o DCNN, which can lead o a long aining ime (Zhang e al., 2021). A he same ime, DCNNs need a sea numbe o expe imen al da a o aining because a small numbe o expe imen al da a may lead o o e i ing p oblems (Gó iz e al., 2020; Zhang e al., 2020). To cope wi h he p oblems men ioned abo e, we p opose h ee no el models o classi y b ain diseases au oma ically. They a e: DenseNe -based Schmid neu al ne wo k (DSNN), DenseNe -based andom ec o unc ional link (DRVFL), and DenseNe -based ex eme lea ning machine (DELM). We selec DenseNe o ex ac ea u es and use andomized neu al ne wo ks (RNNs) o classi ica ion. We modi y he p e- ained DenseNe . Then, he modi ied DenseNe is ine- uned on he da ase . In he DSNN, he las i e laye s wi hin he ine- uned DenseNe a e subs i u ed by he Schmid neu al ne wo k (SNN). In he DRVFL, we selec he RVFL (RVFL) o subs i u e he las i e laye s o he F on ie s in Sys ems Neu oscience | www. on ie sin.o g 3May 2022 | Volume 16 | A icle 838822 Zhu e al. DSNN o B ain Disease Classi ica ion FIGURE 1 | (A) Unheal hy and (B) heal hy b ain images in he da ase . ine- uned DenseNe . In he DELM, we choose he ex eme lea ning machine (ELM) o eplace he end i e laye s o he ine- uned DenseNe .Fi e- old c oss- alida ion is used o e alua e he p oposed h ee models: DSNN, DRVFL, and DELM, in e ms o aspec s (Acc, Sen, Spe, P e, and F1). We inally ge ha DSNN gi es he bes pe o mance among he h ee p oposed models and o e pe o ms he o he six s a e-o - he-a algo i hms. The i e main inno a ions o his s udy a e: (1) DenseNe is alida ed as he backbone by expe imen s showing i s supe io i y o AlexNe , ResNe -18, ResNe -50, and VGG. (2) DSNN, DRVFL, and DELM a e p oposed by eplacing he las i e laye s wi hin he ine- uned DenseNe wi h h ee andomized neu al ne wo ks. (3) The DSNN ge s he bes pe o mance among he h ee p oposed models. (4) The DSNN o e pe o ms he es ic ed DenseNe and spiking neu al ne wo k by expe imen s. (5) The DSNN is compa ed wi h six s a e-o - he-a algo i hms and ob ains he bes esul s among he lis me hods. The es o his a icle is as ollows. The da ase is gi en in Sec ion ‘‘Ma e ials". Sec ion ‘‘Me hodology" discusses he me hodology. Sec ion ‘‘Resul s and Discussion" is abou he expe imen esul s. We conclude his a icle in Sec ion ‘‘Conclusion". MATERIALS The da ase is downloaded om he Ha a d Medical School websi e (Johnson and Becke , 2021). The e a e ou ypes o b ain diseases: ce eb o ascula disease, neoplas ic disease, degene a i e disease, and in lamma o y o in ec ious disease. This a icle classi ies all ou b ain disease images as unheal hy b ain images. A o al o 177 unheal hy b ain images and 20 heal hy b ain images a e used in his a icle. The size o all images in his a icle is 256 ×256. Some unheal hy and heal hy b ain images in his a icle a e shown in Figu e 1. The le ou images a e he unheal hy b ain images, and he igh ou a e he heal hy images. METHODOLOGY P oposed DSNN Tables 2,3gi e he ac onym de ini ions and pa ame e de ini ions, espec i ely. Mo e and mo e esea che s de o e ene gy o esea ching image classi ica ion echnology (Lu S. e al., TABLE 2 | Ac onym and ull explana ion. Ac onym Full explana ion AD Alzheime ’s disease Acc Accu acy A A e age BN Ba ch no maliza ion CNN Con olu ion neu al ne wo k DCNN Deep con olu ion neu al ne wo k DELM DenseNe -based ex eme lea ning machine DL Deep lea ning DRVFL DenseNe -based andom ec o unc ional link DSNN DenseNe -based Schmid neu al ne wo k ELM Ex eme lea ning machine F1 F1-sco e FC Fully connec ed ML Machine lea ning P e P ecision RVFL Random ec o unc ional link RNNs Randomized neu al ne wo ks Sen Sensi i i y SNN Schmid neu al ne wo k Spe Speci ici y S d S anda d de ia ion F on ie s in Sys ems Neu oscience | www. on ie sin.o g 4May 2022 | Volume 16 | A icle 838822 Zhu e al. DSNN o B ain Disease Classi ica ion TABLE 3 | The de ini ion o he pa ame e . Pa ame e De ini ion OmThe ou pu o he M- h laye TmThe nonlinea ans o ma ion (xi,yi)The gi en da ase nThe inpu dimension MThe ou pu dimension wjThe weigh s ec o djThe bias o he j- h hidden node PThe inal ou pu weigh s qThe ou pu biases o SNN Y=(y1, ...., yN)TThe g ound- u h label ma ix o he da ase X=(x1, ...., xN)TThe inpu ma ix s() The sigmoid unc ion VThe numbe o hidden nodes AThe ou pu ma ix o he hidden laye 2021). In image classi ica ion, ea u e ex ac ion is a c ucial s ep. Howe e , he image con ains oo much messy in o ma ion, so ex ac ing aluable ea u es is di icul . Decades ago, people usually manually ex ac ed ea u es. Howe e , manual ea u e ex ac ion akes much ime, and he esul s a e usually no ideal. Wi h he con inuous p og ess o compu e echnology, mo e and mo e people use compu e models o image ea u e ex ac ion (Leming e al., 2020). Many compu e models a e success ul (Lu S. Y. e al., 2021), such as CNN models. The con olu ion laye in he CNN model can signi ican ly educe he olume o pa ame e s o sho en he aining ime. Resea che s ha e p oposed many g ea CNN models, such as AlexNe (Lu e al., 2020a), MobileNe (Lu S.-Y. e al., 2020), ResNe (Lu e al., 2020b), and so on. This a icle p oposes h ee models o he au oma ic classi ica ion o b ain diseases: DSNN, DRVFL, and DELM. The DSNN ge s he bes pe o mance among he h ee p oposed models. The pseudocode o he p oposed DSNN is shown in Table 4. The pipeline o ou model is gi en in Figu e 2. We choose he p e- ained DenseNe as he backbone o he p oposed DSNN. We modi y he p e- ained DenseNe . Then, he modi ied DenseNe is ine- uned on he da ase . The las i e laye s wi hin he ine- uned DenseNe a e subs i u ed by he Schmid neu al ne wo k (SNN). In ou model, he ine- uned DenseNe plays he ole o ea u e ex ac ion. The SNN is ained by he ex ac ed ea u es F om he ine- uned DenseNe . Fi e- old c oss- alida ion is used o e alua e he p oposed DSNN. Backbone o he P oposed DSNN The CNN models (Albawi e al., 2017) ha e been esea ched con inuously in ecen decades. In 1998, LeCun p oposed LeNe (LeCun, 2015) wi h a i e-laye s uc u e. In 2014, he isual geome y g oup p oposed VGG (Simonyan and Zisse man, 2014) wi h a 19-laye s uc u e. The Highway Ne wo ks (S i as a a e al., 2015) we e p oposed la e , wi h mo e han 100 laye s. Wi h he inc easing numbe o ne wo k laye s in CNN models, esea che s a e oubled by he p oblem o g adien anishing. Ba ch no maliza ion (BN) alle ia es he p oblem o g adien anishing o some ex en . ResNe (He e al., 2016) educes he g adien anishing p oblem by cons uc ing iden i y TABLE 4 | Pseudocode o he p oposed DSNN. S ep 1: Load he p e- ained DenseNe . S ep 2: Modi y he p e- ained DenseNe . S ep 2.1 Remo e so max and classi ica ion laye om he p e- ained DenseNe . S ep 2.2 Add FC128, ReLU, BN, FC2, so max, and classi ica ion laye . S ep 3: Di ide he da ase in o i e g oups o he same size and se i=1 S ep 4: Use he i- h g oup as he es se , and all he o he g oups o m he aining se . S ep 5: Fine- une he modi ied DenseNe . S ep 5.1: Inpu is he aining se . S ep 5.2: Ta ge is he co esponding label. S ep 6: Replace he las i e laye s o he ine- uned DenseNe wi h SNN. S ep 7: Ex ac ea u es Fas he ou pu o he FC128 laye . S ep 8: T ain he classi ie o he DSNN on he ex ac ed ea u es Fand he labels. S ep 8.1: Inpu is he ex ac ed ea u es. S ep 8.2: The a ge is he label o he aining se . S ep 8.3: SNN is he classi ie o he DSNN. S ep 9: Tes he ained DSNN on he es se . S ep 10: Repo he es classi ica ion pe o mance o he ained DSNN. S ep 11: Se i=i+1, i i<6, go o S ep 4. S ep 12: A e age es classi ica ion pe o mance. mapping. In 2017, DenseNe (Huang e al., 2017) was p oposed o educe he g adien anishing p oblem by es ablishing dense connec i i y be ween he on and ea laye s. Dense connec i i y makes mo e e ec i e use o ea u es han o he ne wo ks. Thus, DenseNe can achie e be e pe o mance. The gene al iew o DenseNe is gi en in Figu e 3A. Dense blocks e e o he speci ic blocks o DenseNe , as shown in Figu e 3A. All he on laye s a e connec ed wi h he ea laye s. In he same dense block, he heigh and wid h o each ea u e map will no change, bu he numbe o channels will change. In he adi ional sequen ial CNN, i you ha e M laye s, he e will be Mconnec ions, bu DenseNe will in oduce M(M+1)/2 mo e connec ions. Supposing he e a e Mlaye s, OM deno es he ou pu o he M- h laye , TM ep esen s he nonlinea ans o ma ion. The compa ison o DenseNe wi h o he CNNs is lis ed below: T adi ional sequen ial CNN: OM=TM(OM−1)(1) ResNe : OM=TM(OM−1)+OM−1(2) DenseNe : OM=TM([O0,O1, ..., OM−1])(3) whe e [] is he conca ena ion. The ansi ion laye is a module ha connec s di e en dense blocks. I s p ima y unc ion is o in eg a e he ea u es ob ained om he p e ious dense block and educe i s wid h and heigh . Resea che s used he ImageNe da ase o p e- ain he DenseNe . The e a e 1,000 ou pu nodes on he p e- ained DenseNe . Howe e , his a icle only needs wo ou pu nodes. We modi y he p e- ained DenseNe . The modi ica ions a e shown in Figu e 3B. A e hese modi ica ions, we ine- une he modi ied DenseNe by he aining se . We emo e he las i e laye s o he F on ie s in Sys ems Neu oscience | www. on ie sin.o g 5May 2022 | Volume 16 | A icle 838822 Zhu e al. DSNN o B ain Disease Classi ica ion FIGURE 2 | The pipeline o he p oposed DSNN. ine- uned DenseNe and add SNN o imp o e he classi ica ion pe o mance. In he p oposed DSNN, he ine- uned DenseNe is he ea u e ex ac ion. Th ee P oposed Ne wo ks Compa ed wi h he p e- ained DenseNe , andomized neu al ne wo ks (RNNs) ha e a much sho e aining ime. In he DSNN, we eplace he end i e laye s o he ine- uned DenseNe wi h he RNN: he Schmid neu al ne wo k (SNN; Schmid e al., 1992). The SNN is ained by ex ac ed ea u es n om FC128. The s uc u e o he SNN is shown in Figu e 4. The yellow box is he inpu , he pink ci cle ep esen s he hidden nodes, and he g een box shows he ou pu . Gi en N samples and da ase wi h he i- h sample as (xi,yi): xi=(xi1, ..., xin)T∈Rn,i=1, ..., N, (4) yi=(yi1, ..., yim)T∈Rm,i=1, ..., N, (5) whe e nis he inpu dimension, mis he ou pu dimension. The aining algo i hm o SNN is as ollows. The weigh s ec o (wj) connec s he j- h hidden node wi h inpu nodes, dj is he bias o he j- h hidden node. The weigh s ec o (wj) and he bias (dj) a e assigned wi h andom alues and will emain unchanged du ing he aining p ocess. The ou pu ma ix o he hidden laye wi h Vhidden nodes is calcula ed as ollows: ASNN = V X j= 1 s(wjxi+dj),i=1, ..., N, (6) whe e he sigmoid unc ion is ep esen ed as s(). Then we use pseudo-in e se o calcula e he inal ou pu weigh s (P): (P,q)=A† SNNY, (7) whe e he ou pu biases o SNN a e q,A† SNN is he pseudo-in e se ma ix o ASNN, and Y= (y1,...,yN)Tis he g ound- u h label ma ix o he da ase . We p opose wo o he models: DRVFL and DELM. The backbone o he wo o he p oposed models is he p e- ained DenseNe . We modi y he p e- ained DenseNe in he wo p oposed models as he ‘‘modi ica ions o he p e- ained DenseNe ’’ in he DSNN. We eplace he so max and classi ica ion laye o he p e- ained DenseNe wi h six laye s: FC128, ReLU, BN, FC2, so max, and classi ica ion laye . We ine- une he modi ied DensNe by he aining se . In he DRVFL, we selec RVFL (Pao e al., 1994) o subs i u e he las i e laye s o he ine- uned DenseNe . The s uc u e o RVFL is F on ie s in Sys ems Neu oscience | www. on ie sin.o g 6May 2022 | Volume 16 | A icle 838822 Zhu e al. DSNN o B ain Disease Classi ica ion FIGURE 3 | Backbone o he p oposed DSNN. (A) The gene al iew o DenseNe . (B) The modi ica ions in he p e- ained DenseNe . FIGURE 4 | S uc u e o SNN. shown in Figu e 5A. In he DELM, ELM (Huang e al., 2006) is chosen o eplace he las i e laye s o he ine- uned DenseNe . The s uc u e o ELM is shown in Figu e 5B. ELM and RVFL a e wo ypes o RNNs. The yellow box ep esen s he inpu , he pink ci cle deno es he hidden nodes, and he g een box shows he ou pu . The di e ence be ween hese wo RNNs is ha he e a e sho cu connec ions om he inpu o he ou pu in RVFL. The calcula ion s eps a e simila : Gi en Nsamples and da ase wi h he i- h sample as (xi,yi) xi=(xi1, ..., xin)T∈Rn,i=1, ..., N, (8) yi=(yi1, ..., yim)T∈Rm,i=1, ..., N, (9) whe e nis he inpu dimension, mis he ou pu dimension. The aining s eps o hese wo RNNs a e as ollows: S ep 1: wjis he weigh ec o , which connec s he inpu nodes wi h he j- h hidden node. The bias o he j- h hidden node is ep esen ed as dj. We andomly assign wjand djwi h alues. These alues will no change in aining. S ep 2: The hidden laye ’s ou pu ma ix is calcula ed as: Fo RVFL: ARVFL =conca (X,K), (10) whe e X= (x1,...,xN)Tdeno es he inpu ma ix. The Kis calcula ed as ollows: KRVFL = V X j= 1 s(wjxi+dj),i=1, ..., N, (11) whe e Vis he numbe o he hidden nodes in he hidden laye , s() ep esen s he sigmoid unc ion. F on ie s in Sys ems Neu oscience | www. on ie sin.o g 7May 2022 | Volume 16 | A icle 838822 Zhu e al. DSNN o B ain Disease Classi ica ion FIGURE 5 | The s uc u es o (A) RVFL and (B) ELM. Fo ELM: AELM = V X j= 1 s(wjxi+dj),i=1, ..., N, (12) S ep 3: The ou pu weigh s (p): can be calcula ed by pseudo- in e se: Fo RVFL: p=A† RVFLY, (13) whe e A† RVFL is he pseudo-in e se ma ix o ARVFL, and Y= (y1,...,yN)Tis he g ound- u h label ma ix o he da ase . Fo ELM: p=A† ELMY, (14) whe e A† ELM is he pseudo-in e se ma ix o AELM. The backbone o hese o he wo p oposed models in his a icle is he same. The di e ence is ha DRVFL chooses RVFL as i s classi ie , and DELM selec s ELM as i s classi ie . E alua ion We de ine he unheal hy b ain as he posi i e and he heal hy b ain as he nega i e. Fi e indica o s a e chosen o e i y ou model: accu acy (Acc), sensi i i y (Sen), speci ici y (Spe), p ecision (P e), and F1-sco e (F1), espec i ely. Thei o mulas a e shown below:              Acc =TP + TN TP+TN+FP+FN Sen =TP TP + FN Spe =TN TN + FP P e =TP TP + FP F1 =2×TP 2TP + FP + FN (15) whe e he de ini ions o TP, FN, FP, and TN a e he ue posi i e, alse nega i e, alse posi i e, and ue nega i e, espec i ely. TABLE 5 | The hype -pa ame e se ings o he p oposed DSNN. Hype -pa ame e Value Mini-ba ch size 10 Max-epoch 4 Lea ning a e 10−4 Numbe o he hidden nodes V 400 RESULTS AND DISCUSSIONS Expe imen Se ings We modi y he hype -pa ame e se ings o he p oposed DSNN. The max-epoch is se o 4 o educing o e i ing p oblems. We se ou mini-ba ch size o 10 because he da ase is ela i ely small. Acco ding o he expe ience, he lea ning a e is 10−4. A hype -pa ame e we se in ou model is he numbe o hidden nodes (V), which is se as 400 based on he inpu dimension. The hype -pa ame e se ings o ou model a e shown in Table 5. Pe o mances o he DSNN We use i e- old c oss- alida ion o e alua e he p oposed DSNN. The classi ica ion pe o mance o ou model is gi en in Table 6. The Acc, Sen, Spe, P e, and F1 o he p oposed DSNN a e 98.46% ±2.05% , 100.00% ±0.00% , 85.00% ±20.00% , 98.36% ±2.17%, and 99.16% ±1.11% , espec i ely. The esul s o DSNN a e highe han 85%. Especially he sensi i i y is 100%. The ROC cu e is shown in Figu e 6. The AUC alue is 0.9786. I is an e ec i e classi ie when he AUC alue is g ea e han 0.95. These esul s can be concluded ha DSNN is an e ec i e model o classi y b ain diseases. Compa ison o Th ee P oposed Models The classi ica ion pe o mances o DRVFL and DELM based on he i e- old c oss- alida ion a e shown in Table 6. Fo a F on ie s in Sys ems Neu oscience | www. on ie sin.o g 8May 2022 | Volume 16 | A icle 838822 Zhu e al. DSNN o B ain Disease Classi ica ion TABLE 6 | The classi ica ion pe o mance based on i e- old c oss- alida ion (uni : %). Me hods Fold Acc Sen Spe P e F1 DSNN(Ou s) F 1 100.00 100.00 100.00 100.00 100.00 F 2 100.00 100.00 100.00 100.00 100.00 F 3 94.87 100.00 50.00 94.59 97.22 F 4 100.00 100.00 100.00 100.00 100.00 F 5 97.44 100.00 75.00 97.22 98.59 A 98.46 100.00 85.00 98.36 99.16 S d ±2.05 ±0.00 ±20.00 ±2.17 ±1.11 DRVFL(Ou s) F 1 100.00 100.00 100.00 100.00 100.00 F 2 100.00 100.00 100.00 100.00 100.00 F 3 89.74 100.00 0.00 89.74 94.59 F 4 100.00 100.00 100.00 100.00 100.00 F 5 97.44 100.00 75.00 97.22 98.59 A 97.44 100.00 75.00 97.39 98.64 S d ±3.97 ±0.00 ±38.73 ±3.97 ±2.10 DELM(Ou s) F 1 100.00 100.00 100.00 100.00 100.00 F 2 100.00 100.00 100.00 100.00 100.00 F 3 92.31 100.00 25.00 92.11 95.89 F 4 100.00 100.00 100.00 100.00 100.00 F 5 97.44 100.00 75.00 97.22 98.59 A 97.95 100.00 80.00 97.87 98.90 S d ±2.99 ±0.00 ±29.15 ±3.07 ±1.60 Fine- uned DenseNe F 1 87.50 86.11 100.00 100.00 92.54 F 2 82.05 80.00 100.00 100.00 88.89 F 3 89.74 88.57 100.00 100.00 93.94 F 4 85.00 83.33 100.00 100.00 90.91 F 5 79.49 77.14 100.00 100.00 87.10 A 84.76 83.03 100.00 100.00 90.67 S d ±3.67 ±4.10 ±0.00 ±0.00 ±2.46 AlexNe -SNN F 1 89.74 100.00 0.00 89.74 94.59 F 2 89.74 100.00 0.00 89.74 94.59 F 3 90.00 97.22 25.00 92.11 94.59 F 4 90.00 97.22 25.00 92.11 94.59 F 5 89.74 97.14 25.00 91.89 94.44 A 89.84 98.32 15.00 91.12 94.56 S d ±0.13 ±1.38 ±12.25 ±1.13 ±0.06 ResNe -18-SNN F 1 100.00 100.00 100.00 100.00 100.00 F 2 97.50 100.00 75.00 97.30 98.63 F 3 100.00 100.00 100.00 100.00 100.00 F 4 94.87 100.00 50.00 94.59 97.22 F 5 94.87 97.14 75.00 97.14 97.14 A 97.45 99.43 80.00 97.81 98.60 S d ±2.29 ±1.14 ±18.71 ±2.03 ±1.26 ResNe -50-SNN F 1 95.00 94.44 100.00 100.00 97.14 F 2 100.00 100.00 100.00 100.00 100.00 F 3 97.44 97.14 100.00 100.00 98.55 F 4 95.00 100.00 50.00 94.74 97.30 F 5 100.00 100.00 100.00 100.00 100.00 A 97.49 98.32 90.00 98.95 98.60 S d ±2.24 ±2.23 ±20.00 ±2.10 ±1.24 VGG-SNN F 1 97.50 100.00 75.00 97.30 98.63 F 2 87.50 94.44 25.00 91.89 93.15 F 3 94.87 97.14 75.00 97.14 97.14 F 4 89.74 100.00 0.00 89.74 94.59 F 5 87.18 88.57 75.00 96.88 92.54 A 91.36 96.03 50.00 94.59 95.21 S d ±4.12 ±4.26 ±31.62 ±3.16 ±2.33 Res ic ed DenseNe -SNN F 1 94.87 94.29 100.00 100.00 97.06 F 2 100.00 100.00 100.00 100.00 100.00 F 3 97.37 97.06 100.00 100.00 98.51 F 4 94.87 100.00 50.00 94.59 97.22 F 5 100.00 100.00 100.00 100.00 100.00 A 97.42 98.27 90.00 98.92 98.56 S d ±2.09 ±1.83 ±1.97 ±2.09 ±1.69 The bold alues a e esul s o ou p oposed model. F on ie s in Sys ems Neu oscience | www. on ie sin.o g 9May 2022 | Volume 16 | A icle 838822