Role of foundation models in data-driven tissue diagnostics

{"Artificial":[0],"intelligence":[1],"(AI)-based":[2],"tissue":[3],"diagnostics":[4,78],"is":[5,47,167,174,185],"entering":[6],"a":[7,116,162],"new":[8,60],"phase":[9],"driven":[10],"by":[11,169,176],"pathology":[12],"foundation":[13,91,229],"models:":[14],"large-scale":[15],"encoders":[16],"and":[17,23,39,71,93,107,130,138,149,181,193,201,204,217,226,242],"multimodal":[18],"systems":[19],"pretrained":[20],"with":[21,248],"self-supervised":[22],"vision-language":[24],"objectives.":[25],"Yet":[26],"the":[27,34,95],"evidence":[28,158,225],"base":[29],"has":[30],"expanded":[31],"faster":[32],"than":[33,161],"methods":[35],"used":[36],"to":[37,110,236],"evaluate":[38],"interpret":[40],"claimed":[41],"capabilities.":[42],"The":[43],"central":[44],"clinical":[45,80,111],"question":[46],"no":[48],"longer":[49],"whether":[50,72],"these":[51],"models":[52,92,230],"can":[53,74],"achieve":[54],"strong":[55],"retrospective":[56],"performance,":[57],"but":[58],"what":[59,65],"capabilities":[61,68,123],"they":[62,73],"provide,":[63],"under":[64],"conditions":[66],"those":[67],"are":[69,231],"demonstrated,":[70],"translate":[75],"into":[76,255],"reliable":[77,257],"in":[79],"settings.":[81],"This":[82],"review":[83,222],"makes":[84],"three":[85],"contributions.":[86],"First,":[87],"we":[88,114,152,221],"define":[89],"histopathology-centric":[90],"distill":[94],"technical":[96],"factors":[97],"that":[98,120,228],"shape":[99],"their":[100],"behavior:":[101],"pretraining":[102],"data":[103],"regime,":[104],"learning":[105],"objective,":[106],"downstream":[108],"adaptation":[109,199,213],"endpoints.":[112],"Second,":[113],"introduce":[115],"practical":[117],"capability":[118,166,253],"framing":[119],"distinguishes":[121],"general-purpose":[122],"(coverage":[124],"across":[125],"tissues,":[126],"scales,":[127],"stains,":[128],"scanners,":[129],"institutions)":[131],"from":[132],"functional":[133],"breadth":[134],"(task":[135],"primitives,":[136],"adaptability,":[137],"analysis":[139],"level),":[140],"while":[141],"accounting":[142],"for":[143,251],"modality":[144],"scope":[145],"spanning":[146],"vision,":[147],"language,":[148],"genomics.":[150],"Third,":[151],"synthesize":[153],"reported":[154,170],"results":[155],"as":[156],"an":[157],"map":[159],"rather":[160],"leaderboard,":[163],"clarifying":[164],"where":[165,172,182],"supported":[168],"evidence,":[171],"reproducibility":[173],"constrained":[175],"access":[177],"or":[178],"reporting":[179],"limits,":[180],"further":[183],"validation":[184],"needed.":[186],"We":[187,246],"then":[188],"analyze":[189],"current":[190],"benchmarking":[191],"practice":[192],"identify":[194],"common":[195],"confounders,":[196],"including":[197],"heterogeneous":[198],"protocols":[200],"pretraining-evaluation":[202],"overlap,":[203],"propose":[205],"deployment-aware":[206],"recommendations":[207,250],"built":[208],"around":[209],"broad-and-deep":[210],"benchmarks,":[211],"standardized":[212],"\\"budgets,\\"":[214],"overlap":[215],"auditing,":[216],"domain-agnostic":[218],"evaluation.":[219],"Finally,":[220],"emerging":[223],"clinical-utility":[224],"argue":[227],"most":[232],"compelling":[233],"when":[234],"tied":[235],"workflow-defined":[237],"endpoints,":[238],"calibrated":[239],"operating":[240],"points,":[241],"measurable":[243],"operational":[244],"benefit.":[245],"conclude":[247],"actionable":[249],"converting":[252],"demonstrations":[254],"clinically":[256],"data-driven":[258],"diagnostic":[259],"systems.":[260]}