Barzilay, Regina

professor at MIT; researcher in natural language processing and applications of deep learning to oncology and chemistry

Regina Barzilay: Deep Learning for Cancer Diagnosis and Treatment

Lex Fridman Podcast · Sep 23, 2019

• Regina Barzilay
• machine learning for oncology
• breast cancer detection
• medical data access and privacy
• breast density heuristic
• AI drug discovery and molecular graphs
• natural language processing history
• Turing test and chatbots
• few-shot learning
• cancer diagnosis and treatment

Core ideas

Barzilay describes how her 2014 breast cancer diagnosis reshaped her sense of scientific priorities: during the two-and-a-half-month period before she knew her prognosis, she confronted her own mortality concretely for the first time, and returning to MIT afterward she found herself newly disturbed by how much academic energy goes into "trivialities" (incremental metric improvements, narrow linguistic curiosities) relative to the enormous, visible suffering she witnessed daily during treatment. This experience redirected her research toward applying machine learning directly to cancer detection and treatment, while also giving her a personal, humbling window into how little control patients often have — being flooded with contradictory test results, lacking real answers, and encountering close friends and even her own reflexive behavior (mistaking a stranger's slow speech for a disability rather than presumed language difficulty, an experience she connects to a passage in the novel "Americanah") that illustrated how differently the world looks from inside versus outside a crisis.

She frames a key contrast between computer science's practical, prediction-first culture and biology/medicine's traditional emphasis on full mechanistic understanding: recommendation systems like Amazon's don't claim to "understand" users, they simply exploit statistical patterns, and she argues an analogous probabilistic, pattern-matching approach — without necessarily achieving deep causal understanding — can meaningfully advance early cancer diagnosis even where full mechanistic biological understanding remains elusive. She's direct that machine learning's clearest current value in oncology is early detection rather than treatment discovery: cancers like pancreatic cancer are typically diagnosed only once they're already incurable, whereas current statistical risk models (mostly built on family history) are so simplistic that ~80% of breast cancer patients, like Barzilay herself, have no family history and were never flagged as elevated risk — a gap she argues deep learning models, trained on large volumes of imaging and multi-modal data, can close by detecting weak combined signals invisible to the human eye or simplistic heuristics.

Barzilay identifies data access, not algorithmic capability, as oncology AI's central bottleneck: it took her two years to obtain data access after deciding to work in the field, and unlike ImageNet-style benchmarks, there is no publicly available, modern, representative dataset of mammograms in the U.S. — the only broadly available resource being a 1990s film-mammogram dataset too outdated to produce models that generalize to current imaging. She traces this to a structural mismatch of incentives: hospitals bear significant legal and reputational risk in releasing patient data but see comparatively little direct benefit from doing so, and she argues for an "organ donor"-style consent model letting patients proactively opt in to donate their own medical data for research, alongside technical de-identification and encoded/federated learning techniques (some developed at MIT) that let models train on data without institutions ever exposing raw identifiable records. She notes acute personal demand for this kind of infrastructure — she regularly receives mammograms and requests for statistical analysis directly from women facing urgent, high-stakes treatment decisions with insufficient information — and cites the closure of earlier consumer health-record platforms (Google Health, Microsoft HealthVault) as a discouraging sign that this "obvious," Musk-style unsolved infrastructure problem still lacks a working, adopted solution.

She uses breast density assessment as a detailed case study of how slow institutional adoption, not algorithmic capability, gates real-world impact: a 1967 radiologist's qualitative pattern (how much "white" appears on a mammogram) was gradually formalized, and only became U.S. federal law in 2019 requiring density disclosure to patients — despite the fact that dense breasts affect roughly 40-50% of women, making the resulting "high risk" flag nearly useless as an individually actionable signal, and despite deep learning models trained directly on mammogram images now substantially outperforming this decades-old heuristic at predicting actual cancer risk. She argues the real bottleneck to replacing it isn't algorithmic quality (already "orders of magnitude better") but the sociological and regulatory work of convincing hospitals, clinicians, and patients to trust and adopt a new standard — citing Elisabeth Rosenthal's book "An American Sickness" as an eye-opening account of the U.S. healthcare system's complex incentive structures that computer scientists entering the field need to understand to drive real adoption, and suggesting that public, consumer-driven demand (since everyone is a current or future patient) may be the most viable lever for accelerating change, more than top-down policy or institutional advocacy alone.

Separately, Barzilay highlights AI-driven drug discovery as a comparatively unexplored but promising frontier: current pharmaceutical development still relies almost entirely on human chemists' deep domain expertise (running high-throughput lab screens on hundreds of thousands of compounds, then manually iterating on promising "hits" to optimize solubility, toxicity, and side effects) rather than any machine-learning-designed drug, and she describes her own group's work treating molecules as graphs (atoms as nodes, bonds as edges), first learning to predict molecular properties from graph embeddings, then working toward generating novel candidate molecules directly — analogizing this generative process to machine translation, translating an existing molecule into an improved one via a learned latent representation.

On natural language processing, Barzilay traces her own career from 1997, when the field was transitioning from hand-built, linguistically-driven rule systems toward corpus-based statistical methods (she recalls her first paper containing no formulas and no formal evaluation, just example outputs — then the field's norm), through to today's neural systems, noting that linguistic theory's role in NLP research has sharply diminished as statistical and now deep learning methods have come to dominate. She's explicitly outcome-focused rather than concerned with whether machine understanding of language mirrors human cognition — comparing a translation system to a calculator, which computes answers via processes utterly unlike human arithmetic yet remains useful — while acknowledging real concern about the brittleness of current NLP systems under distributional shift (citing a running classroom example: a translation of a Finnish-Karelian pie recipe that has produced unusable garbage output, unchanged, for years, evidently because such text falls far outside the model's effective training distribution). Reflecting on the Turing test, she recounts the story of Xiaoice-style chatbots retaining large, devoted user bases despite trivial underlying mechanisms, and MIT's own ELIZA — whose creator, Joseph Weizenbaum, was reportedly so disturbed by secretaries' willingness to believe in and confide in the simple pattern-matching program that he abandoned the research direction — to argue that passing a Turing-style test may depend less on how sophisticated the technology actually is than on how readily humans are willing to believe it understands them.

Barzilay identifies few-shot and transfer learning (the ability to generalize from very few examples to new tasks or domains, rather than requiring large domain-specific retraining) as the field's most important current gap, arguing most published "improvement from few examples" results remain far from practically usable. Asked about human-level or superhuman AI, she reframes the question around functional breadth rather than mimicking human cognition exactly, and separately raises the possibility of technology directly augmenting human cognition (citing "Flowers for Algernon" and her own experience with a fitness tracker's gamified point system unexpectedly and powerfully shaping her own behavior) as a complementary path to increasing intelligence — alongside direct brain-computer interfaces — while noting simple behavioral-feedback tools already demonstrate how effectively measurement and nudges can reshape human action, for better or worse.

Excerpts